boundary element method for solving inverse heat …boundary element method for solving inverse heat...
TRANSCRIPT
Boundary Element Method for Solving
Inverse Heat Source Problems
Areena Hazanee
Submitted in accordance with the requirements for the degree of
Doctor of Philosophy
The University of Leeds
Department of Applied Mathematics
July 2015
The candidate confirms that the work submitted is his/her own, ex-
cept where work which has formed part of jointly authored publications
has been included. The contribution of the candidate and theother au-
thors to this work has been explicitly indicated below. The candidate
confirms that appropriate credit has been given within the thesis where
reference has been made to the work of others.
This copy has been supplied on the understanding that it is copyright
material and that no quotation from the thesis may be published without
proper acknowledgement.
@2015 The University of Leeds and Areena Hazanee
ii
Joint Publications
Some material of this thesis is based on five published papersand three conference
proceedings as well as one paper pending, as follows:
Publications
1. Hazanee, A., Ismailov, M. I., Lesnic, D. and Kerimov, N. B.(2013) An inverse
time-dependent source problem for the heat equation,Applied Numerical Math-
ematics, Vol.69, pp.13–33.
2. Hazanee, A. and Lesnic, D. (2013) Determination of a time-dependent heat
source from nonlocal boundary conditions,Engineering Analysis with Bound-
ary Elements, Vol.37, No.6, pp.936–956.
3. Hazanee, A. and Lesnic, D. (2013) Reconstruction of an additive space- and
time-dependent heat source,European Journal of Computational Mechanics,
Vol.22, Nos.5–6, pp.304–329.
4. Hazanee, A. and Lesnic, D. (2014) Determination of a time-dependent coef-
ficient in the bioheat equation,International Journal of Mechanical Sciences,
Vol.88, pp.259–266.
5. Hazanee, A., Ismailov, M. I., Lesnic, D. and Kerimov, N. B.(2015) An inverse
time-dependent source problem for the heat equation with a non-classical bound-
ary condition,Applied Mathematical Modelling, in press (available online).
Conference Proceedings
1. Hazanee, A. and Lesnic, D. (2013) The boundary element method for solving an
inverse time-dependent source problem,Proceedings of the 9th UK Conference
on Boundary Integral Methods, (ed. O. Menshykov), UniPrint, University of
Aberdeen, pp.55–62.
iii
2. Hazanee, A. and Lesnic, D. (2014) A time-dependent coefficient identification
problem for the bioheat equation,ICIPE2014 8th International Conference on
Inverse Problems in Engineering, (eds. I. Szczygiel, A. J. Nowak and M. Ro-
jczyk), Silesian University of Technology, May 12-15, 2014, Poland, pp.127–
136.
3. Hazanee, A. and Lesnic, D. (2015) The boundary element method for an in-
verse time-dependent source problem for the heat equation with a non-classical
boundary condition,Proceedings of the 10th UK Conference on Boundary Inte-
gral Methods, (ed. P. J. Harris), University of Brighton, pp.28–35.
Possible Publications
1. Hazanee, A. and Lesnic, D. (2015) Reconstruction of multiplicative space- and
time-dependent sources, submitted to Inverse Problems in Science and Engineer-
ing.
iv
Acknowledgements v
Acknowledgements
In the name of Allah, the most gracious and merciful, Alhamdulillah I thank Al-
lah (subhanahu wa ta’ala) for endowing me with health, patience, and knowledge to
complete this work. This thesis appears in its current form due to the assistance and
guidance of several people. I would like to extend my sincereappreciation to all of
them.
First and foremost, I would like to express my deepest gratitude to my supervisor,
Professor Daniel Lesnic, for providing me with this great opportunity to do my PhD
under his great supervision. My gratitude extends to Dr. EvyKersale, my adviser, for
his valuable advice during my whole PhD study. I also would like to extend my thanks
to all PhD students in the inverse problem group, Bandar Bin-Mohsin, Mohammed
Hussein, Shilan Hussein, Taysir Dyhoum and Dimitra Flouri.
I would like to give my special thanks to all staff in the School of Mathematics
for their help. Many thanks to all my friends in this school, Abeer Al-Nahdi, Wafa
Al-Mohri, Mufida Mohamed Hmaida, and especially to my best friend and sister, Aolo
Bashar Abusaksaka, who was always besides and supported me in any situation I have
faced. I am grateful to Thai students, Joy Riyapan, Jeab Pattamawadee, Jiab Dussadee
for their very good suggestions and providing me the Thai food which made me felt
like I am still in Thailand. And also I am thankful to all Thai Muslim students in the
UK for their friendship and fraternity during my time in the UK.
Of course, the indispensable thanks go to my Hazanee family,dad, mom, brothers,
and my beloved aunt for unconditional love, support, cheering me up and standing by
me. Far distance and time difference could not make me feel I am away from them.
I also gratefully acknowledge the funding support providedby the Ministry of Sci-
ence and Technology of Thailand, Prince of Songkla University Thailand, for pursuing
my PhD at the University of Leeds. I would also like to give special mention to all
staff in the Office of Educational Affairs (OEA) in London fortheir advice and help
dealing with my enquiries.
Abstract vi
Abstract
In this thesis, the boundary element method (BEM) is appliedfor solving inverse
source problems for the heat equation. Through the employment of the Green’s for-
mula and fundamental solution, the BEM naturally reduces the dimensionality of the
problem by one although domain integrals are still present due to the initial condition
and the heat source. We mainly consider the identification oftime-dependent source
for heat equation with several types of conditions such as non-local, non-classical,
periodic, fixed point, time-average and integral which are considered as boundary or
overdetermination conditions. Moreover, the more challenging cases of finding the
space- and time-dependent heat source functions for additive and multiplicative cases
are also considered.
Under the above additional conditions a unique solution is known to exist, however,
the inverse problems are still ill-posed since small errorsin the input measurements re-
sult in large errors in the output heat source solution. Thensome type of regularisation
method is required to stabilise the solution. We utilise regularisation methods such as
the Tikhonov regularisation with order zero, one, two, or the truncated singular value
decomposition (TSVD) together with various choices of the regularisation parameter.
The numerical results obtained from several benchmark testexamples are presented
in order to verify the efficiency of adopted computational methodology. The retrieved
numerical solutions are compared with their analytical solutions, if available, or with
the corresponding direct numerical solution, otherwise. Accurate and stable numeri-
cal solutions have been obtained throughout for all the inverse heat source problems
considered.
Nomenclature vii
Nomenclature
Roman Symbols
A, A0, AL, A(1)k ,A(1)
0,k, A(1)L,k, A
I , AIIi
BEM coefficient matrices
A0j(x, t), ALj(x, t) BEM coefficient functions
B, B∗, B0,BL,B(1)k , B(1)∗
k , B(1)L,k, B
I , BIIi , BIII
BEM coefficient matrices
B0j(x, t), BLj(x, t) BEM coefficient integral functions
C, C(1)k , CI , CII
i , CIII BEM coefficient matrices
Ck([0, T ]) the space ofk-order continuously differentiable functions
on [0, T ]
Ck,l(DT ) the space ofk-order andl-order continuously differentiable
functions in time and space, respectively
d(x, t), d0(x, t), d1(x, t), d2(x, t)
BEM integral functions for the source term
d¯, d
¯I , d
¯II BEM coefficient vectors
D,D0,D1,D2,D(1)k ,DI
1,DI2,DII
1 ,DII2 ,Dr,Dr
k,DrI ,DrIII
BEM coefficient matrices
DT := (0, L)× (0, T ) solution domain
DT := [0, L]× [0, T ] closure of the solution domainDT
ei discrete components of functionE(t)
E(t) mass or energy
E¯
vector ofE(t)
E¯ǫ noisy vector ofE(t)
E = diag(ei) diagonal matrix with componentsei
f(x, t), F (x, t), f(x, t) source functions
F, F0, Fλ objective functions
G(x, t, ξ, τ) fundamental solution for one-dimensional heat equation
Nomenclature viii
h(x, t) given source function
h0j , hLj boundary temperatures
h¯, h
¯0, h
¯Lboundary temperature vectors
H(t) Heaviside step function
Hγ/2[0, T ],Hγ[0, L],H2+γ,1+γ/2(DT )
Holder spaces in Chapter 6
i, j, k indices
k(t), ki(t) given boundary and overdetermination functions
k∗ the largest index of spacex satisfyingxk∗ < X0
k¯
vector of functionk(t)
L length of one-dimensional space domain
n outward unit normal to the space boundary
N number of time steps
N0 number of space cells
Nt truncation number (for TSVD)
p, p0 percentages of perturbation
P (t) perfusion coefficient function
q¯, q
¯0, q
¯Lboundary heat flux vectors
q0j , qLj boundary heat fluxes
r(t), r1(t) time-dependent source functions
R, R0, R1, R2, R(1), R(2), Rregularisation matrices
s(x), s1(x) space-dependent source functions
S coefficient matrix
S0, S0 values of the sources at the fixed point
T , T1, T2 final and fixed times
u(x, t), u(x, t) temperatures
u0(x), u0(x) initial temperature
u0,k initial temperature atxk
Nomenclature ix
u¯0
, u¯0
initial temperature vectors
U¯, V
¯orthogonal vectors for SVD
v(x, t) transformation function
v1(t), v2(t) given functions
V1, V2 diagonal matrices ofv1(t), v2(t)
W12(0, T ), W
22(0, T ), W
22(0, L), W
42(0, L), W
4,22 (DT )
Sobolev space in Chapter 7
w¯
unknown vector ofr(t) ands(x)
x, y, x spaces (variable)
X,Xi left-hand side coefficient matrices of BEM linear system
X0 fixed location
Xǫ contaminated left-hand side coefficient matrix
yn eigenfunction of the spectral problem
y¯, y
¯iright-hand side vector of BEM linear system
y¯ǫ noisy right-hand side vector
Greek Symbols
α, a, b heat transfer coefficients
β(x), β1(x), β2(x), β(x, t)
given functions
δij the Kronecker delta symbol
ψ(x) given function
χ(t), χ1(t) given functions
ǫ, ǫ0 noise levels
γ1, γ2 fixed point source values
γij(t) given functions
λ, λL, λdis, λGCV , λopt regularisation parameters
µ1(t), µ2(t), µ0(t), µL(t)
given boundary temperatures
Contents x
Ω domain
∂Ω boundary of the domainΩ
Ω = Ω ∪ ∂Ω closure of the domainΩ
Φ4n0
space of functions
τ , t times (variable)
σ, σψ, σχ, σµ, σr, σs standard deviations
σi singular values
Σ diagonal matrix with components ofσi
ϑ(x, t) given function
Abbreviations
BEM boundary element method
CBEM constant boundary element method
erf, erfc error and complementary error functions
FDM finite difference method
FEM finite element method
FOTR first-order Tikhonov regularisation
GCV generalised cross-validation
PDE partial differential equation
RMSE root mean square error
SOTR second-order Tikhonov regularisation
SVD singular value decomposition
TSVD truncated singular value decomposition
ZOTR zeroth-order Tikhonov regularisation
Contents
Joint Publications iii
Acknowledgements v
Abstract vi
Nomenclature vii
Contents xi
List of Figures xiv
List of Tables xxvi
1 General Introduction 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Inverse and ill-posed problems . . . . . . . . . . . . . . . . . . . . .1
1.3 The boundary element method (BEM) . . . . . . . . . . . . . . . . . 2
1.4 The BEM for solving one-dimensional direct heat problem. . . . . . 4
1.5 Condition number . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.6 Regularisations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.6.1 The truncated singular value decomposition (TSVD) . .. . . 9
1.6.2 The Tikhonov regularisation . . . . . . . . . . . . . . . . . . 10
1.6.3 Choice of the regularisation parameter . . . . . . . . . . . .. 11
xi
Contents xii
1.7 Purpose of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2 Determination of a Time-dependent Heat Source from Nonlocal Boundary
Conditions 17
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 Mathematical formulation . . . . . . . . . . . . . . . . . . . . . . . 18
2.3 The boundary element method (BEM) . . . . . . . . . . . . . . . . . 22
2.3.1 Case 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3.2 Case 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3.3 Case 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.3.4 Case 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3.5 Case 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3.6 Case 6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.4 Numerical examples and discussion . . . . . . . . . . . . . . . . . .31
2.4.1 Example 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4.2 Example 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3 Determination of a Time-dependent Heat Source from Nonlocal Boundary
and Integral Conditions 61
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.2 Mathematical formulation . . . . . . . . . . . . . . . . . . . . . . . 62
3.3 The boundary element method (BEM) . . . . . . . . . . . . . . . . . 63
3.4 Regularisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.5 Numerical examples and discussion . . . . . . . . . . . . . . . . . .67
3.5.1 Example 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.5.2 Example 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Contents xiii
4 Determination of a Time-dependent Coefficient in the Bioheat Equation 83
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.2 Mathematical formulation . . . . . . . . . . . . . . . . . . . . . . . 84
4.3 The boundary element method (BEM) . . . . . . . . . . . . . . . . . 85
4.4 Regularisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.5 Numerical examples and discussion . . . . . . . . . . . . . . . . . .89
4.5.1 Example 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.5.2 Example 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5 Determination of a Time-dependent Heat Source with a Dynamic Bound-
ary Condition 99
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.2 Mathematical formulation . . . . . . . . . . . . . . . . . . . . . . . 100
5.3 Boundary element method (BEM) . . . . . . . . . . . . . . . . . . . 103
5.4 Numerical examples and discussion . . . . . . . . . . . . . . . . . .106
5.4.1 Example 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.4.2 Example 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.4.3 Example 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
6 Determination of an Additive Space- and Time-dependent Heat Source 121
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
6.2 Mathematical formulation . . . . . . . . . . . . . . . . . . . . . . . 122
6.3 The boundary element method (BEM) . . . . . . . . . . . . . . . . . 125
6.4 Regularisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.5 Numerical examples and discussion . . . . . . . . . . . . . . . . . .131
6.5.1 Example 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
6.5.2 Example 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
6.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
List of Figures xiv
7 Determination of Multiplicative Space- and Time-dependent Heat Sources149
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
7.2 Mathematical formulation . . . . . . . . . . . . . . . . . . . . . . . 150
7.3 The boundary element method (BEM) . . . . . . . . . . . . . . . . . 152
7.4 Solution of inverse problem . . . . . . . . . . . . . . . . . . . . . . . 155
7.5 Numerical examples and discussion . . . . . . . . . . . . . . . . . .156
7.5.1 Example 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
7.5.2 Example 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
7.5.3 Example 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
7.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
8 General Conclusions and Future Work 177
8.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
8.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
Bibliography 185
List of Figures
2.1 The normalised singular values of matrixX for (a) Case 1 – (f) Case 6
for N0 = N = 20 (− · −), 40 (· · · ), 80 (−−−), for Example 1. . . 33
2.2 The analytical (—–) and numerical (− · −) results of (a)r(t), (b)
u(0, t), and (c)u(1, t) for exact data andλ = 0, for Example 1 Case 1. 35
2.3 The analytical (—–) and numerical (− · −) results ofr(t) obtained by
using the zeroth-order Tikhonov regularisation with the regularisation
parameters (a)λ = 10−7 gives RMSE=0.319, and (b)λ = 10−5 gives
RMSE=0.499, for exact data for Example 1 Case 1. . . . . . . . . . . 36
2.4 The analytical (—–) and numerical (− · −) results of (a)r(t), (b)
u(0, t), and (c)u(1, t) for p = 1% noisy data andλ = 0, for Example
1 Case 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.5 (a) The discrepancy principle curve, and the analytical(—–) and nu-
merical results of (b)r(t), (c) u(0, t), and (d)u(1, t) obtained using
the zeroth-order Tikhonov regularisation forp = 1% (− · −) and
p = 3% (− − −) noise with the regularisation parametersλdis given
in Table 2.2, for Example 1 Case 1. . . . . . . . . . . . . . . . . . . . 38
2.6 (a) The discrepancy principle curve, and the analytical(—–) and nu-
merical results of (b)r(t), (c) u(0, t), and (d)u(1, t) obtained us-
ing the first-order Tikhonov regularisation forp = 1% (− · −) and
p = 3% (− − −) noise with the regularisation parametersλdis given
in Table 2.2, for Example 1 Case 1. . . . . . . . . . . . . . . . . . . . 39
xv
List of Figures xvi
2.7 The analytical (—–) and numerical (− · −) results of (a)r(t), (b)
u(1, t), and (c)ux(0, t) for exact data andλ = 0, for Example 1 Case 2. 40
2.8 The analytical (—–) and numerical (− · −) results of (a)r(t), (b)
u(1, t), and (c)ux(0, t) for p = 1% noisy data andλ = 0, for Ex-
ample 1 Case 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.9 (a) The discrepancy principle curve, and the analytical(—–) and nu-
merical results of (b)r(t), (c) u(1, t), and (d)ux(0, t) obtained us-
ing the zeroth-order Tikhonov regularisation forp = 1% (− · −) and
p = 3% (− − −) noise with the regularisation parametersλdis given
in Table 2.3, for Example 1 Case 2. . . . . . . . . . . . . . . . . . . . 42
2.10 (a) The discrepancy principle curve, and the analytical (—–) and nu-
merical results of (b)r(t), (c) u(1, t), and (d)ux(0, t) obtained us-
ing the first-order Tikhonov regularisation forp = 1% (− · −) and
p = 3% (− − −) noise with the regularisation parametersλdis given
in Table 2.3, for Example 1 Case 2. . . . . . . . . . . . . . . . . . . . 43
2.11 The analytical (—–) and numerical (− · −) results of (a)r(t), (b)
ux(0, t), and (c)ux(1, t) for p = 1% noisy data andλ = 0, for Ex-
ample 1 Case 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.12 (a) The discrepancy principle curve, and the analytical (—–) and nu-
merical results of (b)r(t), (c) ux(0, t), and (d)ux(1, t) obtained us-
ing the zeroth-order Tikhonov regularisation forp = 1% (− · −) and
p = 3% (− − −) noise with the regularisation parametersλdis given
in Table 2.4, for Example 1 Case 3. . . . . . . . . . . . . . . . . . . . 45
2.13 (a) The discrepancy principle curve, and the analytical (—–) and nu-
merical results of (b)r(t), (c) ux(0, t), and (d)ux(1, t) obtained us-
ing the first-order Tikhonov regularisation forp = 1% (− · −) and
p = 3% (− − −) noise with the regularisation parametersλdis given
in Table 2.4, for Example 1 Case 3. . . . . . . . . . . . . . . . . . . . 46
List of Figures xvii
2.14 The analytical (—–) and numerical (− · −) results of (a)r(t), (b)
u(1, t), (c) ux(0, t), andux(1, t) for p = 1% noisy data andλ = 0,
for Example 1 Case 4. . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.15 (a) The discrepancy principle curve, and the analytical (—–) and nu-
merical results of (b)r(t), (c) u(1, t), (d) ux(0, t), and (e)ux(1, t) ob-
tained using the zeroth-order Tikhonov regularisation forp = 1% (− ·−) andp = 3% (−−−) noise with the regularisation parametersλdis
given in Table 2.5, for Example 1 Case 4. . . . . . . . . . . . . . . . 49
2.16 (a) The discrepancy principle curve, and the analytical (—–) and nu-
merical results of (b)r(t), (c) u(1, t), (d) ux(0, t), and (e)ux(1, t) ob-
tained using the first-order Tikhonov regularisation forp = 1% (−·−)
andp = 3% (− − −) noise with the regularisation parametersλdis
given in Table 2.5, for Example 1 Case 4. . . . . . . . . . . . . . . . 50
2.17 The analytical (—–) and numerical (− · −) results of (a)r(t), (b)
u(0, t), (c) u(1, t) and (d)ux(1, t) for exact data andλ = 0, for Exam-
ple 1 Case 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.18 The analytical (—–) and numerical (− · −) results of (a)r(t), (b)
u(0, t), (c) u(1, t) and (d)ux(1, t) for p = 1% noisy data andλ = 0,
for Example 1 Case 5. . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.19 (a) The discrepancy principle curve, and the analytical (—–) and nu-
merical results of (b)r(t), (c) u(0, t), (d) u(1, t), and (e)ux(1, t) ob-
tained using the zeroth-order Tikhonov regularisation forp = 1% (− ·−) andp = 3% (−−−) noise with the regularisation parametersλdis
given in Table 2.6, for Example 1 Case 5. . . . . . . . . . . . . . . . 53
2.20 (a) The discrepancy principle curve, and the analytical (—–) and nu-
merical results of (b)r(t), (c) u(0, t), (d) u(1, t), and (e)ux(1, t) ob-
tained using the first-order Tikhonov regularisation forp = 1% (−·−)
andp = 3% (− − −) noise with the regularisation parametersλdis
given in Table 2.6, for Example 1 Case 5. . . . . . . . . . . . . . . . 54
List of Figures xviii
2.21 The analytical (—–) and numerical (− · −) results of (a)r(t), (b)
u(0, t), (c) u(1, t), (d) ux(0, t), and (e)ux(1, t) for exact data and
λ = 0, for Example 1 Case 6. . . . . . . . . . . . . . . . . . . . . . . 55
2.22 (a) The discrepancy principle curve, and the analytical (—–) and nu-
merical results of (b)r(t), (c) u(0, t), (d) u(1, t), (e)ux(0, t), and (f)
ux(1, t) obtained using the zeroth-order Tikhonov regularisation for
p ∈ 1(− · −), 3(· · · ), 5(− − −)% noise with the regularisation
parametersλdis given in Table 2.7, for Example 1 Case 6. . . . . . . . 56
2.23 (a) The discrepancy principle curve, and the analytical (—–) and nu-
merical results of (b)r(t), (c) u(0, t), (d) u(1, t), (e)ux(0, t), and (f)
ux(1, t) obtained using the first-order Tikhonov regularisation forp ∈1(− · −), 3(· · · ), 5(−−−)% noise with the regularisation param-
etersλdis given in Table 2.7, for Example 1 Case 6. . . . . . . . . . . 57
2.24 (a) The discrepancy principle curve, and the analytical (—–) and nu-
merical results of (b)r(t) obtained using the first-order Tikhonov reg-
ularisation forp ∈ 0(), 0.1(−·−), 0.5(· · · ), 1(−−−)% noise
with the regularisation parametersλdis given in Table 2.8, for Example 2. 58
3.1 The normalised singular values of matrixX for N0 = N ∈ 20 (− ·−), 40 (· · · ), 80 (−−−), for Example 1. . . . . . . . . . . . . . . . 68
3.2 The analytical (—–) and numerical (− · −) results of (a)r(t) and (b)
u(0, t) for exact data andλ = 0, for Example 1. . . . . . . . . . . . . 69
3.3 The analytical (—–) and numerical (− · −) results of (a)r(t) and (b)
u(0, t) for p = 1% noisy data andλ = 0, for Example 1. . . . . . . . 70
3.4 TheL-curve on (a), (c), (e) log-log scale and (b), (d), (f) normal
scale, obtained using TSVD forp ∈ 1(top), 3(middle), 5(bottom)%noise, for Example 1. . . . . . . . . . . . . . . . . . . . . . . . . . . 72
List of Figures xix
3.5 (a) The discrepancy principle curve, and the analytical(—–) and nu-
merical results of (b)r(t), and (c)u(0, t) obtained using the TSVD for
p ∈ 1(− · −), 3(· · · ), 5(− − −)% noise at the truncation levelNt
given in Table 3.1, for Example 1. . . . . . . . . . . . . . . . . . . . 73
3.6 (a) The discrepancy principle curve, and the analytical(—–) and nu-
merical results of (b)r(t), and (c)u(0, t) obtained using the ZOTR for
p ∈ 1(− · −), 3(· · · ), 5(− − −)% noise with the regularisation
parametersλdis given in Table 3.1, for Example 1. . . . . . . . . . . . 74
3.7 (a) The discrepancy principle curve, and the analytical(—–) and nu-
merical results of (b)r(t), and (c)u(0, t) obtained using the FOTR for
p ∈ 1(− · −), 3(· · · ), 5(− − −)% noise with the regularisation
parametersλdis given in Table 3.1, for Example 1. . . . . . . . . . . . 76
3.8 (a) The discrepancy principle curve, and the analytical(—–) and nu-
merical results of (b)r(t), and (c)u(0, t) obtained using the SOTR for
p ∈ 1(− · −), 3(· · · ), 5(− − −)% noise with the regularisation
parametersλdis given in Table 3.1, for Example 1. . . . . . . . . . . . 77
3.9 The numerical results of (a)E(t) and (b)u(0, t) obtained by solving
the direct problem withN0 = N ∈ 20 (−·−), 40 (· · · ), 80 (−−−),
for Example 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.10 The analytical (—–) and numerical(− · −) results of (a)r(t) and (b)
u(0, t) for exact data andλ = 0, for Example 2. . . . . . . . . . . . . 78
3.11 (a) The discrepancy principle curve, and the analytical (—–) and nu-
merical results of (b)r(t), and (c)u(0, t) obtained using the ZOTR for
p ∈ 1(− · −), 3(· · · ), 5(− − −)% noise with the regularisation
parametersλdis given in Table 3.2, for Example 2. . . . . . . . . . . . 80
3.12 (a) The discrepancy principle curve, and the analytical (—–) and nu-
merical results of (b)r(t), and (c)u(0, t) obtained using the FOTR for
p ∈ 1(− · −), 3(· · · ), 5(− − −)% noise with the regularisation
parametersλdis given in Table 3.2, for Example 2. . . . . . . . . . . . 81
List of Figures xx
3.13 (a) The discrepancy principle curve, and the analytical (—–) and nu-
merical results of (b)r(t), and (c)u(0, t) obtained using the SOTR for
p ∈ 1(− · −), 3(· · · ), 5(− − −)% noise with the regularisation
parametersλdis given in Table 3.2, for Example 2. . . . . . . . . . . . 82
4.1 The analytical (—–) and numerical(− · −) results of (a)r(t), (b)
u(0, t), (c) r′(t), and (d)P (t) obtained using no regularisation, i.e.
λ = 0, for exact data, for Example 1. . . . . . . . . . . . . . . . . . . 90
4.2 The analytical (—–) and numerical(− · −) results of (a)r(t), (b)
u(0, t), (c) r′(t), and (d)P (t) obtained using no regularisation, i.e.
λ = 0, for p = 1% noise, for Example 1. . . . . . . . . . . . . . . . . 91
4.3 The GCV function (4.21), obtained using the second-order Tikhonov
regularisation forp = 1% noise, for Example 1. . . . . . . . . . . . . 92
4.4 The analytical (—–) and numerical results of (a)r(t), (b) u(0, t), (c)
r′(t), and (d)P (t) obtained using the second-order Tikhonov regular-
isation withλGCV = 1.25 × 10−5 (− · −) andλopt = 1.05 × 10−6
( ), for p = 1% noise, for Example 1. . . . . . . . . . . . . . . . 93
4.5 The analytical (—–) and numerical results of (a)r(t), (b) u(0, t)(t),
(c) r′(t), and (d)P (t) obtained using no regularisation in (4.31), i.e.
Λ = 0, for exact data( ) and forp = 1% noisy data(− · −), for
Example 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.6 The analytical (—–) and numerical results of (a)r′(t) and (b)P (t)
obtained using the smoothing spline regularisation withΛdis = 7 ×10−5 ( ) andΛ = 8 × 10−3 (− · −), as defined in (4.34), for
p = 1% noise, for Example 2. . . . . . . . . . . . . . . . . . . . . . . 97
5.1 The analytical (—–) and numerical results(− · −) of (a) r(t), (b)
u(1, t), (c) ux(0, t), and (d)ux(1, t) for exact data, for Example 1. . . 108
List of Figures xxi
5.2 The analytical (—–) and numerical results of (a)r(t), (b) u(1, t), (c)
ux(0, t), and (d)ux(1, t) obtained using the straightforward inversion
(− · −) with no regularisation, and the second-order Tikhonov regu-
larisation() with the regularisation parameterλ=4.3E-6 suggested
by the GCV method, forp = 1% noise, for Example 1. . . . . . . . . 109
5.3 The analytical (—–) and numerical results of (a)r(t), (b) u(1, t), (c)
ux(0, t), and (d)ux(1, t) obtained using the second-order Tikhonov
regularisation with the regularisation parameter suggested by the GCV
method, forp = 3% (· · ·) andp = 5% (−−−), for Example 1. . . . 110
5.4 The numerical results of (a)u(1, t), (b) ux(0, t), (c) ux(1, t), and (d)
E(t) obtained by solving the direct problem withN = N0 ∈ 20( ), 40(· · ·), 80(−−−), for Example 2. . . . . . . . . . . . . . . . 112
5.5 The analytical solution (5.21) and the direct problem numerical solu-
tion from Figures 5.7(a)–5.7(c) (—–) and numerical resultsof (a)r(t),
(b) u(1, t), (c) ux(0, t), and (d)ux(1, t), with no regularisation, for ex-
act data( ) and noisy datap = 1% (− · −), for Example 2. . . . . 114
5.6 The analytical solution (5.21) and the direct problem numerical solu-
tions from Figures 5.7(a)–5.7(c) (—–), and the numerical results of
(a) r(t), (b) u(1, t), (c) ux(0, t), and (d)ux(1, t) obtained using the
second-order Tikhonov regularisation with the regularisation parame-
ters suggested by GCV method, forp ∈ 1(), 3(· · · ), 5(−−−)%noise, for Example 2. . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.7 The numerical results of (a)u(1, t), (b) ux(0, t), (c) ux(1, t), and (d)
E(t) obtained by solving the direct problem withN = N0 ∈ 20 (− ·−), 40 (· · ·), 80 (−−−), for Example 3. . . . . . . . . . . . . . . 117
5.8 The analytical solution (5.4.3) and the direct problem numerical so-
lution from Figures 5.7(a)–5.7(c) (—–), and numerical results of (a)
r(t), (b) u(1, t), (c) ux(0, t), and (d)ux(1, t), with no regularisation,
for exact data( ) and noisy datap = 1% (− · −), for Example 3. 118
List of Figures xxii
5.9 The analytical solution (5.4.3) and the direct problem numerical solu-
tions from Figures 5.7(a)–5.7(c) (—–), and the numerical results of (a)
r(t), (b)u(1, t), (c)ux(0, t), and (d)ux(1, t) obtained using the second-
order Tikhonov regularisation with the regularisation parameters sug-
gested by the discrepancy method, forp ∈ 1(− · −), 3(· · · ), 5(−−−)% noise, for Example 3. . . . . . . . . . . . . . . . . . . . . . . 119
6.1 The normalised singular values of matrixX forN = N0 ∈ 20, 40, 80andX0 ∈ 1
4(− · −), 1
2(· · · ), 3
4(−−−), for Example 1. . . . . . . 132
6.2 The analytical (—–) and numerical results(−·−) of (a)r(t), (b) s(x),
(c) ux(0, t), and (d)ux(1, t) obtained using the SVD for exact data, for
Example 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.3 (a) TheL-curve and (b) the GCV function obtained by the TSVD for
exact data, for Example 1. . . . . . . . . . . . . . . . . . . . . . . . 135
6.4 (a) TheL-curve, and (b) the GCV function, obtained by the ZOTR
(− · −), FOTR (· · · ), and SOTR(− − −) for exact data, withλ =
λr = λs, for Example 1. . . . . . . . . . . . . . . . . . . . . . . . . . 135
6.5 The analytical (—–) and numerical results of (a)r(t), (b) s(x), (c)
ux(0, t), and (d)ux(1, t) obtained using the TSVD(− + −), ZOTR
(−·−), FOTR(· · · ), and SOTR(−−−) with regularisation parameters
suggested by the GCV function of Figure 6.3(b) and 6.4(b) forexact
data, for Example 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
6.6 The analytical (—–) and numerical results(−·−) of (a)r(t), (b) s(x),
(c) ux(0, t), and (d)ux(1, t) obtained using the SOTR with the regular-
isation parameterλL=1E-1 suggested by theL-curve of Figure 6.4(a)
for exact data, for Example 1. . . . . . . . . . . . . . . . . . . . . . . 137
6.7 (a) TheL-curve and (b) the discrepancy principle obtained using the
TSVD for noisy inputp = 1%, for Example 1. . . . . . . . . . . . . . 138
List of Figures xxiii
6.8 (a) TheL-curve and (b) the discrepancy principle obtained using the
ZOTR(−·−), FOTR(· · · ), and SOTR(−−−) for noisy inputp = 1%,
with λ = λr = λs, for Example 1. . . . . . . . . . . . . . . . . . . . 138
6.9 The analytical (—–) and numerical results of (a)r(t), (b) s(x), (c)
ux(0, t), and (d)ux(1, t) obtained using the TSVD(−+−) with Nt =
14, and the ZOTR(− · −), FOTR (· · · ), SOTR(− − −) with regu-
larisation parameters suggested by the discrepancy principle of Figure
6.8(b) for noisy inputp = 1%, for Example 1. . . . . . . . . . . . . . 139
6.10 TheL-surface on (a) a three-dimensional plot, (b) plane oflog ‖Xw¯λ
−y¯ǫ‖ versuslog ‖R(1)r
¯λ‖, and (c) plane oflog ‖Xw
¯λ−y
¯ǫ‖ versuslog ‖R(2)s
¯λ‖,
obtained using the SOTR for noisy inputp = 1%, for Example 1. . . . 140
6.11 The analytical (—–) and numerical results of (a)r(t), (b) s(x), (c)
ux(0, t), and (d)ux(1, t) obtained using the SOTR with regularisation
parameters suggested by theL-curve criterionλL = λr = λs = 10
(−−−), theL-surface method(λr,L, λs,L)=(10,1)(−+−), and the trial
and error(λr,opt, λs,opt)=(8,5.2E-2)(− ∗ −), for noisy inputp = 1%,
for Example 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
6.12 The normalised singular values of matrixX forN = N0 = 20 (−·−),
N = N0 = 40 (· · · ), andN = N0 = 80 (−−−), for Example 2. . . . 143
6.13 TheL-curve obtained using (a) the TSVD and (b) the ZOTR(− · −),
FOTR(· · · ), and SOTR(−−−) with λ = λr = λs, for exact data, for
Example 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6.14 The analytical (—–) and numerical results of (a)r(t), (b) s(x), (c)
ux(0, t), and (d)ux(1, t) obtained using the SVD(−·−) and the SOTR
(−o−) with the regularisation parameterλL=1E-4 suggested by theL-
curve of Figure 6.13(b) for exact data, for Example 2. . . . . . .. . . 145
6.15 TheL-curve obtained using (a) the TSVD and (b) the ZOTR(− · −),
FOTR (· · · ), and SOTR(− − −) with λ = λr = λs, for noisy input
p = 1%, for Example 2. . . . . . . . . . . . . . . . . . . . . . . . . . 146
List of Figures xxiv
6.16 TheL-surface on (a) a three-dimensional plot, (b) plane oflog ‖Xw¯λ
−y¯ǫ‖ versuslog ‖R(1)r
¯λ‖, and (c) plane oflog ‖Xw
¯λ−y
¯ǫ‖ versuslog ‖R(2)s
¯λ‖,
obtained using the SOTR for noisy inputp = 1%, for Example 2. . . . 147
6.17 The analytical (—–) and numerical results of (a)r(t), (b) s(x), (c)
ux(0, t), and (d)ux(1, t) obtained using the SOTR with regularisation
parameters suggested by theL-curve criterionλL = λr = λs = 1
(− − −), theL-surface method(λr,L, λs,L)=(1,10)(− + −), and the
trial and error(λr,opt, λs,opt)=(2.2,5.9)(−∗−), for noisy inputp = 1%,
for Example 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
7.1 The analytical (—–) and numerical results for (a)χ(t) and (b)β(x)
obtained using the BEM for the direct problem withN = N0 ∈10 (− · −), 20 (· · · ), 40 (−−−), for Example 1. . . . . . . . . . . 158
7.2 (a) The objective functionF0 and the numerical results for (b)r(t), (c)
s(x), (d) u(0, t), (e)u(0.1, t) obtained with no regularisation(− · −),
for exact data for Example 1. The corresponding analytical solutions
are shown by continuous line (—–) in (b)–(e) and thep0 = 100%
perturbed initial guesses are shown by dotted line(· · · ) in (b) and (c). 160
7.3 The numerical results for (a)r(t), (b) s(x), (c) u(0, t), (d) u(0.1, t)
obtained with the first-order regularisation(· · · ) and the second-order
regularisation(−−−) with regularisation parameterλopt = 10−5, for
exact data for Example 1. The corresponding analytical solutions are
shown by continuous line (—–). . . . . . . . . . . . . . . . . . . . . 161
7.4 (a) TheL-curve criterion, (b) the objective functionFλ, and the nu-
merical results(− −) for (c) r(t), (d) s(x), (e) u(0, t), (f) u(0.1, t)
obtained with the hybrid-order regularisation (7.26) withregularisation
parameterλL = 10−5 suggested byL-curve, for exact data Example 1.
The corresponding analytical solutions are shown by continuous line
(—–) in (c)–(f). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
List of Figures xxv
7.5 (a) The objective functionFλ and the numerical results for (b)r(t), (c)
s(x), (d) u(0, t), (e) u(0.1, t) obtained with the hybrid-order regular-
isation (7.26) with regularisation parameterλL = 10−5 suggested by
L-curve forp ∈ 1(−·−), 3(· · · ), 5(−−−)% noisy data, for Exam-
ple 1. The corresponding analytical solutions are shown by continuous
line (—–) in (b)–(e). . . . . . . . . . . . . . . . . . . . . . . . . . . 164
7.6 The analytical (—–) and numerical results for (a)χ(t) and (b)β(x)
obtained using the BEM for the direct problem withN = N0 ∈ 5(−·−), 10(· · · ), 20(−−−), for Example 2. . . . . . . . . . . . . . . 167
7.7 (a) The objective functionF0, (b) the RMSEs forr(t) (−−−) ands(x)
(· · · ) obtained with no regularisation for exact data, and the numerical
results for (c)r(t) and (d)s(x) obtained using the minimisation pro-
cess after 56 unfixed iterations(−·−), and 31 fixed iterations(), for
Example 2. The corresponding analytical solutions (7.30) are shown
by continuous line (—–) in (c) and (d). . . . . . . . . . . . . . . . . 168
7.8 (a) The objective functionFλ, (b) the RMSEs forr(t) (−−−) ands(x)
(· · · ) obtained using the hybrid-order regularisation (7.26) with regu-
larisation parameterλopt = 2× 10−4 for exact data, and the numerical
results for (c)r(t) and (d)s(x) obtained using minimisation process
after 28 unfixed iterations(− · −), and 23 fixed iterations( ), for
Example 2. The corresponding analytical solutions (7.30) are shown
by continuous line (—–) in (c) and (d). . . . . . . . . . . . . . . . . 169
7.9 (a) The objective functionFλ, (b) the RMSEs forr(t) (− − −) and
s(x) (· · · ) obtained using the hybrid-order regularisation (7.26) with
regularisation parameterλopt = 4 × 10−4 for noise levelǫ0 = 0.01,
and the numerical results for (c)r(t) and (d)s(x) obtained using the
minimisation process after 27 unfixed iterations(− · −), and 21 fixed
iterations( ), for Example 2. The corresponding analytical solu-
tions (7.30) are shown by continuous line (—–) in (c) and (d).. . . . 170
List of Tables xxvi
7.10 (a) The objective functionFλ, (b) the RMSEs forr(t) (− − −) and
s(x) (· · · ) obtained using the hybrid-order regularisation (7.26) with
regularisation parameterλopt = 2 for noise levelǫ0 = 0.1, and the
numerical results(− · −) for (c) r(t) and (d)s(x) obtained using the
minimisation process after 17 (unfixed) iterations, for Example 2. The
corresponding analytical solutions (7.30) are shown by continuous line
(—–) in (c) and (d). . . . . . . . . . . . . . . . . . . . . . . . . . . 171
7.11 The numerical results for (a)χ(t) and (b)β(x) obtained using the BEM
for the direct problem withN = N0 ∈ 10(− · −), 20(· · · ), 40(−−−), for Example 3. . . . . . . . . . . . . . . . . . . . . . . . . . . 173
7.12 (a) The objective functionFλ and the numerical results for (b)r(t) and
(c) s(x) obtained with the hybrid-order regularisation (7.26) withreg-
ularisation parameterλ = 2×10−4 for p ∈ 1(−·−), 5(· · · ), 10(−−−)% noisy data for Example 3. The corresponding analytical solu-
tions (7.33) and (7.34) are shown by continuous line (—–) in (b) and
(c). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
List of Tables
2.1 The condition numbers of the matrixX in equation (2.48) forN =
N0 ∈ 20, 40, 80 for Example 1 Cases 1–6. . . . . . . . . . . . . . . 34
2.2 The RMSE for the zeroth- and first-order Tikhonov regularisation for
p ∈ 0, 1, 3% noise, for Example 1 Case 1. . . . . . . . . . . . . . . 34
2.3 The RMSE for the zeroth- and first-order Tikhonov regularisation for
p ∈ 0, 1, 3% noise, for Example 1 Case 2. . . . . . . . . . . . . . . 42
2.4 The RMSE for the zeroth- and first-order Tikhonov regularisation for
p ∈ 0, 1, 3% noise, for Example 1 Case 3. . . . . . . . . . . . . . . 47
2.5 The RMSE for the zeroth- and first-order Tikhonov regularisation for
p ∈ 0, 1, 3% noise, for Example 1 Case 4. . . . . . . . . . . . . . . 48
2.6 The RMSE for the zeroth- and first-order Tikhonov regularisation for
p ∈ 0, 1, 3% noise, for Example 1 Case 5. . . . . . . . . . . . . . . 48
2.7 The RMSE for the zeroth- and first-order Tikhonov regularisation for
p ∈ 0, 1, 3, 5% noise, for Example 1 Case 6. . . . . . . . . . . . . . 48
2.8 The RMSE for the first-order Tikhonov regularisation forp ∈ 0, 0.1, 0.5, 1%noise, for Example 2. . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.1 The RMSE for the TSVD, ZOTR, FOTR, SOTR forp ∈ 0, 1, 3, 5%noise, for Example 1. . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.2 The RMSE for the ZOTR, FOTR, SOTR forp ∈ 1, 3, 5% noise and
N = 40,N0 = 30, for Example 2. . . . . . . . . . . . . . . . . . . . 79
xxvii
List of Tables xxviii
5.1 The RMSE foru(1, t), ux(0, t), ux(1, t) andE(t) obtained using the
BEM for the direct problem withN = N0 ∈ 20, 40, 80, for Example 1.107
5.2 The regularisation parametersλ and the RMSE forr(t), u(1, t),ux(0, t)
andux(1, t), obtained using the BEM withN = N0 = 40 combined
with the second-order Tikhonov regularisation forp ∈ 0, 1, 3, 5%noise, for Example 1. . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.3 The regularisation parametersλ and the RMSE forr(t), u(1, t),ux(0, t)
andux(1, t), obtained using the BEM withN = 40 andN0 = 30 com-
bined with the second-order Tikhonov regularisation forp ∈ 0, 1, 3, 5%noise, for Example 2. . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.4 The regularisation parametersλ given by the discrepancy principle and
by the GCV, and the RMSE forr(t), u(1, t), ux(0, t) andux(1, t), ob-
tained using the BEM withN = 40 andN0 = 30 combined with the
second-order Tikhonov regularisation forp ∈ 0, 1, 3, 5% noise. The
optimal regularisation parameter given by the minimum of RMSE of
r(t) is also included, for Example 3. . . . . . . . . . . . . . . . . . . 120
6.1 The condition numbers of the matrixX in equation (6.21) for various
N = N0 ∈ 20, 40, 80 andX0 ∈ 14, 12, 34, for Example 1. . . . . . . 133
6.2 The RMSE forr(t), s(x), ux(0, t), andux(1, t) obtained using the
SVD, TSVD, ZOTR, FOTR, and SOTR, forp ∈ 0, 1%, for Example 1.142
6.3 The RMSE forr(t), s(x), ux(0, t), andux(1, t) obtained using the
SVD, TSVD, ZOTR, FOTR, and SOTR, forp ∈ 0, 1%, for Example 2.144
7.1 The RMSE foru(0, t), u(0.1, t), χ(t) andβ(x), obtained using the
BEM for the direct problem withN = N0 ∈ 10, 20, 40, for Example 1.158
7.2 The RMSE forr(t), s(x), u(0, t), u(0.1, t) for exact data, Example 1. 165
7.3 The RMSE forr(t) ands(x), for the noise levelsǫ0 ∈ 0, 0.01, 0.1,
for Example 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
Chapter 1
General Introduction
1.1 Introduction
Inverse problems are becoming an essential part in the development of several appli-
cations in science and engineering such as in medical diagnosis and therapy, ground-
water/air pollution phenomena, or the designing of thermalequipment, systems and
instruments. Such problems, particularly for the heat equation, have important appli-
cations in the field of applied sciences such as in melting andfreezing processes, the
designing and manufacturing areas in which the strength of heat sources is not ex-
actly recognised, especially in the discovery of the quantity of energy generation in a
computer chip, in a microwave heating process, or in a chemical reaction process.
In this thesis, the interest is specialised to solve severalinverse source problems for
the heat equation using the boundary element method (BEM).
1.2 Inverse and ill-posed problems
A direct problem consists of solving a system where an input cause is given and an
output effect is desired. However, if the situation is reversed then we have an inverse
problem which is in general ill-posed (improperly-posed, incorrectly-posed). For more
definitions and examples of inverse and ill-posed problems see the excellent review by
1
Chapter 1. 2
Kabanikhin [31].
The study of ill-posed problems began in the early 20th century through the defi-
nition of well-posedness given by J. Hadamard in 1902. In thesense of Hadamard, a
mathematical problem is well-posed if it satisfies the following properties:
• Existence: For all (suitable) data, there exists a solution of the problem (in an
appropriate sense).
• Uniqueness: For all (suitable) data, the solution is unique.
• Stability: The solution depends continuously on its data (i.e. small perturbations
in the input data do not result in large perturbations in the solution).
According to above definition, any mathematical problem is ill-posed if any one of
these three conditions is violated. In the cases investigated in this thesis, the problems
violate the third condition, i.e. stability.
The main purpose of this thesis focuses on applying BEM to inverse heat source
problems, which are in generally ill-posed in the sense thatsmall measurement errors
greatly magnify the sought solutions.
1.3 The boundary element method (BEM)
One of the main advantage of the BEM over domain discretisation methods such as
the finite-difference method (FDM) or the finite element mehtod (FEM) is that the
discretisation is necessary only on the boundary, i.e. the BEM uses less number of
nodes and elements when compared to the FDM and the FEM. The main idea of the
BEM, which is based on using the Green’s identity and the fundamental solution, is
to find the solution inside the domain by using the solution tothe partial differential
equation (PDE) on the boundary only.
The mathematical background of the BEM is represented by theknowledge of the
fundamental solution and the application of the Green’s identities. We first introduce
Chapter 1. 3
the Heaviside step function and the Dirac delta distribution as follows:
The Heaviside step function:
H(t) =
1, if t > 0,
0, if t ≤ 0.
The Dirac delta distribution function:
δ(x, ξ) = δ(x− ξ) =
0, if x 6= ξ,
∞, if x = ξ.
The fundamental properties of the Dirac delta distributionare
δ(x) = H ′(x),
∫
Ω
f(ξ)δ(x, ξ) dξ = f(x), x ∈ Ω.
Basically, the one-dimensional transient heat equation isgoverned by the partial
differential heat operatorL := ∂2
∂x2− ∂
∂t. Let L > 0 andT > 0 be the length of the
space domain and the time duration, respectively, and definethe solution domain
DT := (0, L)× (0, T ]. (1.1)
Consider the classical heat equation
Lu(x, t) = ∂2u
∂x2(x, t)− ∂u
∂t(x, t) = 0, (x, t) ∈ DT . (1.2)
A functionG(x, t, y, τ) is called a fundamental solution for the heat equation (1.2)if
L∗G(x, t, y, τ) = −δ(x, t; y, τ) = −δ(|x− y|, |t− τ |), (1.3)
whereL∗ = ∂2
∂x2+ ∂
∂tis the adjoint ofL, (x, t) is a field point, and(y, τ) is a source
point. Solving (1.3) using the method of Fourier transform gives the fundamental
Chapter 1. 4
solution, see [60],
G(x, t, y, τ) =H(t− τ)√
4π(t− τ)exp
(
−(x− y)2
4(t− τ)
)
. (1.4)
In order to develop the BEM, let us introduce the Green’s identities, as follows:
∫
Ω
(
U∇2V − V∇2U)
dΩ =
∫
∂Ω
(
U∂V
∂n− V
∂U
∂n
)
dS,
∫
Ω
(
U∇2V +∇U · ∇V)
dΩ =
∫
∂Ω
U∂V
∂ndS,
(1.5)
for any functionsU, V ∈ C2(Ω), wheren is the outward normal to the boundary∂Ω
of the bounded domainΩ.
1.4 The BEM for solving one-dimensional direct heat
problem
In order to understand how the BEM performs, let us consider the direct classical (one-
dimensional) heat conduction problem which requires finding the temperatureu(x, t)
satisfying the heat equation
ut = uxx + F (x, t), (x, t) ∈ DT , (1.6)
whereF is a heat source, subject to the initial condition
u(x, 0) = u0(x), x ∈ [0, L], (1.7)
and the Neumann boundary conditions
ux(0, t) = µ1(t), ux(L, t) = µ2(t), t ∈ (0, T ], (1.8)
Chapter 1. 5
(or the Dirichlet boundary conditions)
u(0, t) = µ1(t), u(L, t) = µ2(t), t ∈ (0, T ]. (1.9)
Mixed or Robin boundary conditions can also be considered. For using the BEM, we
first multiply (1.6) byG and integrate overDT to result in
∫
DT
G(x, t, y, τ)∂u
∂τ(y, τ) dydτ =
∫
DT
G(x, t, y, τ)∂2u
∂y2(y, τ) dydτ
+
∫
DT
G(x, t, y, τ)F (y, τ) dydτ.
Using the Green’s identities (1.5) gives
∫
DT
G(x, t, y, τ)∂u
∂τ(y, τ) dydτ
=
∫ T
0
[
G(x, t, ξ, τ)∂u
∂n(ξ)(ξ, τ)− u(ξ, τ)
∂G
∂n(ξ)(x, t, ξ, τ)
]
ξ∈0,L
dτ
+
∫
DT
u(y, τ)∂2G
∂y2(x, t, y, τ) dydτ +
∫
DT
G(x, t, y, τ)F (y, τ) dydτ, (1.10)
wheren is the outward unit normal to the space boundary0, L, i.e. ∂∂n(ξ)
= − ∂∂ξ
for
ξ = 0, and ∂∂n(ξ)
= ∂∂ξ
for ξ = L. Then, using that the fundamental solution satisfies
(1.3) and the property of the Dirac delta function result in the integral equation
η(x)u(x, t) =
∫ t
0
[
G(x, t, ξ, τ)∂u
∂n(ξ)(ξ, τ)− u(ξ, τ)
∂G
∂n(ξ)(x, t, ξ, τ)
]
ξ∈0,L
dτ
+
∫ L
0
G(x, t, y, 0)u(y, 0) dy+
∫ L
0
∫ T
0
G(x, t, y, τ)F (y, τ) dτdy,
(x, t) ∈ [0, L]× (0, T ], (1.11)
whereη(0) = η(L) = 12
andη(x) = 1 for x ∈ (0, L).
The discretisation of the integral equation (1.11) is performed by dividing the
boundaries0 × (0, T ] andL × (0, T ] into a series ofN small boundary elements
Chapter 1. 6
[tj−1, tj] for j = 1, N , tj =jTN, j = 0, N , whilst the space domain[0, L]× 0 is dis-
cretised into a series ofN0 small cells[xk−1, xk] for k = 1, N0, xk = kN0
, k = 0, N0.
Over each boundary element(tj−1, tj], the temperatureu and the flux∂u∂n
are assumed
to be constant and take their values at the midpointtj =tj−1 + tj
2, i.e.
u(0, t) = u(0, tj) =: h0j , u(L, t) = u(L, tj) =: hLj , t ∈ (tj−1, tj] (1.12)
∂u
∂n(0, t) =
∂u
∂n(0, tj) =: q0j ,
∂u
∂n(L, t) =
∂u
∂n(L, tj) =: qLj , t ∈ (tj−1, tj ]. (1.13)
Note that sincen is the outward unit normal to the (one-dimensional) space boundary,
then
q0j = −∂u∂x
(0, tj), qLj =∂u
∂x(L, tj). (1.14)
In each cell[xk−1, xk], the temperatureu is assumed to be constant and takes its value
at the midpointxk =xk−1 + xk
2, i.e.
u(x, 0) = u(xk, 0) =: u0,k, x ∈ (xk−1, xk]. (1.15)
Also, for the source functionF (x, t), we assume the piecewise constant approximation
in time as
F (x, t) = F (x, tj), t ∈ (tj−1, tj ]. (1.16)
With these approximations, the integral equation (1.11) isdiscretised as
η(x)u(x, t) =
N∑
j=1
[A0j(x, t)q0j + ALj(x, t)qLj −B0j(x, t)h0j −BLj(x, t)hLj ]
+
N0∑
k=1
Ck(x, t)u0,k +N∑
j=1
D0,j(x, t), (x, t) ∈ [0, L]× (0, T ], (1.17)
where the integral coefficients are given by
Aξj(x, t) =
∫ tj
tj−1
G(x, t, ξ, τ)dτ, ξ = 0, L, (1.18)
Chapter 1. 7
Bξj(x, t) =
∫ tj
tj−1
∂G
∂n(ξ)(x, t, ξ, τ)dτ, ξ = 0, L, (1.19)
Ck(x, t) =
∫ xk
xk−1
G(x, t, y, 0)dy, (1.20)
and the double integral source term is given by
D0,j(x, t) =
∫ tj
tj−1
∫ L
0
G(x, t, y, τ)F (y, tj) dydτ. (1.21)
The integrals in expressions (1.18)–(1.20) can be evaluated analytically as, see [15],
Aξj(x, t) =
0 ; t ≤ tj−1,√
t− tj−1
π; tj−1 < t ≤ tj , x = ξ,
|x− ξ|2√π
(
e−z2
0
z0−√
πerfc(z0)
)
; tj−1 < t ≤ tj , x 6= ξ,
√
t− tj−1
π−√
t− tjπ
; t > tj, x = ξ,
|x− ξ|2√π
(
e−z2
0
z0− e−z
2
1
z1+√π (erf(z0)− erf(z1))
)
; t > tj, x 6= ξ,
(1.22)
Bξj(x, t) =
0 ; t ≤ tj−1,
0 ; tj−1 < t ≤ tj, x = ξ,
−erfc(z0)2
; tj−1 < t ≤ tj, x 6= ξ,
erf(z0)− erf(z1)2
; t > tj ,
(1.23)
Ck(x, t) =1
2
[
erf
(
x− xk−1
2√t
)
− erf
(
x− xk
2√t
)]
, (1.24)
whereξ ∈ 0, L, z0 =|x− ξ|
2√t− tj−1
, z1 =|x− ξ|2√t− tj
and erf, erfc are the error func-
tions defined by erf(x) =2√π
∫ x
0
e−σ2
dσ, erfc(x) = 1 − erf(x), respectively. Mean-
Chapter 1. 8
while the double integral (1.21) becomes
D0,j(x, t) =
∫ tj
tj−1
∫ L
0
G(x, t, y, τ)F (y, tj) dydτ =
∫ L
0
F (y, tj)Ayj(x, t) dy,
and can be evaluated using the midpoint rule for numerical integration.
Hence on considering the BEM, we apply the initial condition(1.7) at the nodes
xk for k = 1, N0, as in (1.15), and the integral equation (1.17) at the boundary nodes
(0, ti) and(L, ti) for i = L,N . This gives the system of2N linear equations
Aq¯− Bh
¯+ Cu
¯0+ d
¯= 0
¯, (1.25)
where
A =
A0j(0, ti) ALj(0, ti)
A0j(L, ti) ALj(L, ti)
2N×2N
,
B =
B0j(0, ti) +12δij BLj(0, ti)
B0j(L, ti) BLj(L, ti) +12δij
2N×2N
, C =
Ck(0, ti)
Ck(L, ti)
2N×N0
,
q¯=
q0j
qLj
2N
, h¯=
h0j
hLj
2N
, u¯0
=[
u0,k
]
N0
, d¯=
∑Nj=1D0,j(0, ti)
∑Nj=1D0,j(L, ti)
2N
,
whereδij is the Kronecker delta symbol, defined byδij = 1 for i = j, andδij = 0
for i 6= j. Note that matrix termB also includes the contribution from the left-hand
side of equation (1.17). At this stage, we can find the boundary temperature h¯, if the
Neumann boundary conditions (1.8) are prescribed as
h¯= B−1
(
Aq¯+ Cu
¯0+ d
¯
)
.
Whereas if the Dirichlet boundary conditions (1.9) are prescribed, we can then obtain
the heat flux q¯
as
q¯= A−1 (Bh
¯− Cu
¯0− d
¯) .
Chapter 1. 9
1.5 Condition number
The insight into the degree of conditioning of the system of equations (1.25) is merely
given by the condition number of a matrix herein defined as theratio between the
largest to the smallest singular values. Obviously, the larger the condition number is
the more ill-conditioned is our system of equations.
1.6 Regularisations
Inverse problems are well-known to be in general ill-posed by violating the stability
condition at least. Upon discretisation, this results in anill-conditioned systems of
equations to be solved. To deal with these difficulties the inverse problem is usually
solved as an optimisation problem with regularisation in order to achieve the stability
of the solution. Below we briefly describe two such classicalmethods of regularisation.
1.6.1 The truncated singular value decomposition (TSVD)
Suppose we wish to solve the system ofM linear equations withN unknowns
Xr¯= y
¯ǫ, (1.26)
whereyǫ is a noisy perturbation of the exact right-hand side vector y¯, i.e. ‖y−yǫ‖ ≈ ǫ.
We first decompose the matrixX in the form,
X = UΣV T, (1.27)
whereU = [U¯1,U¯2, . . . ,U¯N
] andV = [V¯1,V¯2, . . . ,V¯N
] areM × N matrices with
columns U¯j
and V¯j
for j = 1, N , such thatUTU = I = V TV , and
Σ = diag(σ1, σ2, . . . , σN) is anN diagonal matrix containing the singular values of
Chapter 1. 10
the matrixX, σj for j = 1, N , in decreasing order
σ1 ≥ σ2 ≥ · · · ≥ σN ≥ 0.
Then the matrix system (1.26) can be reformed to obtain the singular value decompo-
sition (SVD) solution as follows:
r¯=
(
N∑
j=1
1
σjV¯j
· U¯
Tj
)
y¯ǫ. (1.28)
In MATLAB, this decomposition is operated using the command[U,Σ, V ] = svd(X)
or [U,Σ, V ] = svds(X,N). For ill-posed problems, the truncation of (1.28) is needed
to be considered as a regularisation method, by omitting itslastN −Nt small singular
values, whereNt denotes the truncation level. This way, the regularised solution is
given by
r¯Nt
=
(
Nt∑
j=1
1
σjV¯j
· U¯
Tj
)
y¯ǫ, (1.29)
which is simply a truncated SVD (TSVD) stable solution of thefull SVD unstable so-
lution (1.28). And the MATLAB command for the TSVD becomes[UNt,ΣNt
, VNt] =
svds(X,Nt) whereΣNt= diag(σ1, σ2, . . . , σNt
) andUNt=[
U¯1,U¯2, . . . ,U¯NT
]
, VNt=
[
V¯1,V¯2, . . . ,V¯Nt
]
.
1.6.2 The Tikhonov regularisation
Alternatively, the Tikhonov regularisation is another wayof obtaining a stable solution
of the ill-conditioned system of equations (1.26). This method is based on minimising
the regularised linear least-squares objective function,[42, 57],
‖Xr¯− y
¯ǫ‖2 + λ‖Rr‖2 (1.30)
Chapter 1. 11
whereR is a (differential) regularisation matrix of orderk ∈ 0, 1, 2, . . . imposing a
Ck-smoothing constraint on the solution, andλ > 0 is a regularisation parameter to be
prescribed. Note that the norm‖ · ‖ is defined as the Euclidean norm of vector. In this
study, we are considering the order of regularisation matrix R to be order zero, one,
two as defined by [15, 57],
R0 =
1 0 0 .
0 1 0 .
0 0 1 .
. . . .
, the zeroth-order regularisation, (1.31)
R1 =
1 −1 0 0 .
0 1 −1 0 .
0 0 1 −1 .
. . . . .
, the first-order regularisation, (1.32)
R2 =
1 −2 1 0 0 .
0 1 −2 1 0 .
0 0 1 −2 1 .
. . . . . .
, the second-order regularisation. (1.33)
On solving the minimisation of (1.30) one obtains the regularised solution
r¯λ
=(
XTX + λRTR)−1
XTy¯ǫ. (1.34)
1.6.3 Choice of the regularisation parameter
The regularisation parameterλ is very important in (1.34) (also the truncation levelNt
in (1.29)) and it can be chosen according to many criteria, e.g. theL-curve method
[16], the generalised cross-validation (GCV) [63], or the discrepancy principle [40].
TheL-curve method suggests choosingλ at the corner of theL-curve which is a plot
of the norm of the residual‖Xr¯λ−y
¯ǫ‖ versus the solution norm‖r
¯λ‖. Alternatively, the
Chapter 1. 12
discrepancy principle choosesλ > 0 such that the residual‖Xr¯λ
− y¯ǫ‖ ≈ ǫ. Whereas
the GCV criterion suggests choosing the parameterλ as the minimum of the GCV
function,
GCV (λ) =‖Xr
¯λ− y
¯ǫ‖2
[trace(I −X(XTX + λRTR)−1XT)]2, λ > 0. (1.35)
Note that both theL-curve and the GCV criteria are heuristic methods, which arenot
always convert [58], because they do not require the knowledge of the level of noiseǫ.
Then these two methods do not guarantee to give the regularisation parameter.
1.7 Purpose of the thesis
In this thesis, we mainly consider inverse heat source problems for the heat equation
(1.6), whereu is the unknown temperature andF is a heat source term to be identified.
We focus on the identification of the source termF (x, t) in (1.6) in various special
cases. This approach is necessary because otherwise there will be no unique solution
to the inverse problem unlessu(x, t) is specified or measured throughout the whole
solution domainDT , [56].
Moreover, even though uniqueness of solution can be ensuredby restricting the
source term to be of certain special forms, e.g. space-dependent, time-dependent, ad-
ditive or multiplicative, the inverse problem is still ill-posed in the sense that the con-
tinuous dependence upon the input data is violated (small errors in the input data give
rise to large errors in the estimated results). This has to bedealt with by using some
sort of regularisation, e.g. the TSVD as described in Subsection 1.6.1, the Tikhoknov
regularisation as described in Subsection 1.6.2 [1, 15, 63], the iterative algorithm [30],
the variational method [29], the augmented Tikhonov regularisation derived from a
Bayesian perspective [65], the mollification methods [68, 69], the smoothing spline
approximation [59], etc.
The structure of the thesis is as follows. In Chapter 1, the background knowledge
Chapter 1. 13
of inverse and ill-posed problems is provided. The BEM is detailed together with the
application to the classical heat equation.
In Chapter 2, three general boundary conditions of inverse heat source problems are
considered to determine the time-dependent heat sourcer(t) in F (x, t) = r(t)f0(x, t)+
f1(x, t) and the temperatureu(x, t) in the heat equation (1.6), subject to the initial con-
dition (1.7), and the following three general nonlocal boundary and overdetermination
conditions:
γ11(t)u(0, t) + γ12(t)u(L, t) + γ13(t)ux(0, t) + γ14(t)ux(L, t) = k1(t)
γ21(t)u(0, t) + γ22(t)u(L, t) + γ23(t)ux(0, t) + γ24(t)ux(L, t) = k2(t)
γ31(t)u(0, t) + γ32(t)u(L, t) + γ33(t)ux(0, t) + γ34(t)ux(L, t) = k3(t)
, (1.36)
where(ki)i=1,3 are given functions and(γij)i=1,3,j=1,4 is a given matrix of coefficients
having rank 3. The BEM is combined with the Tikhonov regularisation in order to
obtain an accurate and stable numerical solution.
In Chapter 3, we investigate an identification of the time-dependent heat source, i.e.
we seekr(t) in F (x, t) = r(t)f(x, t), together with the temperatureu(x, t) in the heat
equation (1.6), subject to the initial condition (1.7), theperiodic and Robin boundary
conditions
u(0, t) = u(1, t), t ∈ [0, T ], (1.37)
ux(0, t) + αu(0, t) = 0, t ∈ [0, T ], (1.38)
whereα 6= 0 is a given constant, and the integral additional measurement
∫ 1
0
u(x, t)dx = E(t), t ∈ [0, T ]. (1.39)
In this inverse problem, the BEM is developed as a numerical method and combined
with two case studies of the regularisation method. Firstly, we apply the BEM together
with the TSVD method in order to obtain a stable solution, andthen the BEM is con-
Chapter 1. 14
sidered again and combined with various orders of Tikhonov’s regularisation method.
In Chapter 4, we determine the time-dependent blood perfusion coefficient function
P (t) ≥ 0 and the temperatureu(x, t) in the following bioheat equation
ut(x, t) = uxx(x, t)− P (t)u(x, t) + f(x, t), (x, t) ∈ (0, 1)× (0, T ], (1.40)
wheref is a given heat source term. We subject this bioheat equationto the ini-
tial condition (1.7), the boundary conditions (1.37) and (1.38), and the integral over-
determination condition (1.39). A simple transformation is used to reduce the bioheat
equation (1.40) to the classical heat equation (1.6). The BEM for the heat equation
is employed, together with either the second-order Tikhonov regularisation combined
with finite differences, or with a smoothing spline regularisation technique for com-
puting the first-order derivative of a noisy function.
Chapter 5 presents an investigation for the identification of the time-dependent heat
sourcer(t) in F (x, t) = r(t)f(x, t) and the temperatureu(x, t) in the heat equation in
(1.6), subject to the initial condition (1.7), the non-classical boundary condition
auxx(1, t) + αux(1, t) + bu(1, t) = 0, t ∈ [0, T ], (1.41)
wherea, b, α are given numbers not simultaneously equal to zero, and the over-
determination condition (1.39). We are using the same techniques as before.
More challenging, the purpose in Chapter 6 is the simultaneous determination of
an additive space- and time-dependent heat sources, i.e. identifying the unknown com-
ponentsr(t) ands(x) in the source termF (x, t) = r(t)f(x, t) + s(x)g(x, t) + h(x, t),
together with the temperatureu(x, t) in the heat equation (1.6), subject to the initial
condition (1.7), the Dirichlet boundary conditions,
u(0, t) = µ0(t), u(L, t) = µL(t), t ∈ [0, T ], (1.42)
and additional conditions. These latter ones consist of a specified temperature measure-
Chapter 1. 15
ment at an internal pointX0 ∈ (0, L), a time-average temperature, and an additional
fixing conditions, as follows:
u (X0, t) = χ(t), t ∈ [0, T ], (1.43)∫ T
0
u(x, t) dt = ψ(x), x ∈ [0, L], (1.44)
s(X0) = S0. (1.45)
The mathematical problem is linear but ill-posed since the continuous dependence on
the input data is violated. In discretised form the problem reduces to solving an ill-
conditioned system of linear equations. We investigate theperformances of several
regularisation methods, i.e. the TSVD and the Tikhonov regularisation, and examine
their stability with respect to noise in the input data.
A nonlinear heat source problem is finally studied in Chapter7. This consists
of the simultaneous determination of multiplicative space- and time-dependent source
componentsf(t) andg(x) in F (x, t) = r(t)s(x), in the heat equation (1.6), subject to
the initial condition (1.7), the homogeneous Neumann boundary conditions
ux(0, t) = ux(L, t) = 0, t ∈ [0, T ], (1.46)
the specified interior measurement (1.43), the final time temperature measurement at
the ‘upper-base’ final timet = T , and an additional fixing condition, as follows:
u(x, T ) = β(x), x ∈ [0, L], (1.47)
s(X0) = S0. (1.48)
For the numerical discretisation, the BEM combined with a regularised nonlinear op-
timisation are utilised.
Finally, in Chapter 8, the conclusions which summarise the main work of this thesis
and possible future work are highlighted.
Chapter 2. 16
Chapter 2
Determination of a Time-dependent
Heat Source from Nonlocal Boundary
Conditions
2.1 Introduction
Recently, nonlocal boundary and overdetermination conditions have become a centre
of interest in the mathematical formulation and numerical solution of several inverse
and improperly posed problems in transient heat conduction, see e.g. [23, 24, 33, 51],
to mention only a few. They opened a new area of applied numerical and mathematical
modelling research. Practical applications of nonlocal boundary value problems are
encountered in chemical diffusion for heat conduction in biological processes, see e.g.
[11, 41, 46]. For example, in multiphase flows involving fluids, solids and gases, the
heat flux is often taken to be proportional to the difference in boundary temperature
between the various phases, and the quantitiesγij , i = 1, 3, j = 1, 4, present in the
nonlocal boundary condition (2.3) below (see also equation(1.36)) represent those
proportionality factors.
In this chapter, we consider obtaining the numerical solution of several inverse
17
Chapter 2. 18
time-dependent heat source problems for the heat equation with non-local boundary
and overdetermination conditions whose unique solvabilities have previously been in-
vestigated/established by Ivanchov [28]. The mathematical inverse formulations are
described in Section 2.2. Since the inverse problems under investigations are linear,
but ill-posed (in the sense that the continuous dependence upon the input data is vio-
lated), the numerical method is based on the boundary element direct solver combined
with the Tikhonov regularisation, as described in Section 2.3. The choice of the regu-
larisation parameter in the latter procedure is based on thediscrepancy principle, [40].
The above combination yields accurate and stable numericalsolutions, as it will be
presented and discussed in Section 2.4. Finally, Section 2.5 highlights the conclusions
of this chapter.
2.2 Mathematical formulation
Consider the problem of finding the time-dependent heat sourcer(t) ∈ C([0, T ]) and
the temperatureu(x, t) ∈ C2,1(DT ) ∩ C1,0(DT ) which satisfy the heat conduction
equation
ut = uxx + r(t)f(x, t) + h(x, t), (x, t) ∈ DT , (2.1)
subject to the initial condition (1.7), namely
u(x, 0) = u0(x), x ∈ [0, L], (2.2)
and the following general boundary and overdetermination conditions:
γ11(t)u(0, t) + γ12(t)u(L, t) + γ13(t)ux(0, t) + γ14(t)ux(L, t) = k1(t)
γ21(t)u(0, t) + γ22(t)u(L, t) + γ23(t)ux(0, t) + γ24(t)ux(L, t) = k2(t)
γ31(t)u(0, t) + γ32(t)u(L, t) + γ33(t)ux(0, t) + γ34(t)ux(L, t) = k3(t)
(2.3)
Chapter 2. 19
where
f, h ∈ C1,0(DT ), u0 ∈ C2([0, L]), ki ∈ C1([0, T ]), i = 1, 3, (2.4)
and the matrixγ = (γij)i=1,3,j=1,4 ∈ C1([0, T ]) has rank 3 for allt ∈ [0, T ].
Actually, this inverse problem was studied theoretically by Ivanchov [28] who es-
tablished its unique solvability. Moreover, by assuming, without any loss of generality,
that the same third-order minor of the matrixγ is non-zero we can express three of the
four boundary datau(0, t), u(L, t), ux(0, t), andux(L, t) in terms of the fourth one and
distinguish the following six cases:
Case 1 ux(0, t) = µ1(t), ux(L, t) = µ2(t), (2.5)
v1(t)u(0, t) + v2(t)u(L, t) = k(t); (2.6)
Case 2 u(0, t) = µ1(t), ux(L, t) = µ2(t), (2.7)
v1(t)ux(0, t) + v2(t)u(L, t) = k(t); (2.8)
Case 3 u(0, t) = µ1(t), u(L, t) = µ2(t), (2.9)
v1(t)ux(0, t) + v2(t)ux(L, t) = k(t); (2.10)
Case 4 u(0, t) = µ1(t), ux(L, t) + v1(t)u(L, t) = µ2(t), (2.11)
ux(0, t) + v2(t)ux(L, t) = k(t); (2.12)
Case 5 ux(0, t) = µ1(t), ux(L, t) + v1(t)u(L, t) = µ2(t), (2.13)
u(0, t) + v2(t)u(L, t) = k(t); (2.14)
Case 6 ux(0, t)− v1(t)u(0, t) = µ1(t), ux(L, t) + v2(t)u(L, t) = µ2(t), (2.15)
v3(t)u(0, t) + v4(t)u(L, t) = k(t), (2.16)
for t ∈ [0, T ], wherek ∈ C1([0, T ]) is a given function resulted from manipulating the
system (2.3). The other mixed boundary conditions cases corresponding to Cases 2,
4–6 can be reduced to these ones by the change of variabley = L− x.
The following Theorems 2.2.1–2.2.5 from [28] give the unique solvability, i.e. ex-
Chapter 2. 20
istence and uniqueness of the solutions of the inverse problem (2.1)–(2.3) in all the
above six cases.
Theorem 2.2.1 Assume that the regularity conditions(2.4)are satisfied and that:
(i) v1, v2 ∈ C1([0, T ]), v21(t) + v22(t) > 0, t ∈ [0, T ];
(ii) v1(t)f(0, t) + v2(t)f(L, t) 6= 0, t ∈ [0, T ];
(iii) µ1(0) = u′0(0), µ2(0) = u′0(L), v1(0)u0(0) + v2(0)u0(L) = k(0).
Then the inverse problem(2.1), (2.2), (2.5), (2.6)representing Case 1 is uniquely solv-
able.
Theorem 2.2.2 Assume that, in addition to conditions(2.4)and(i) of Theorem 2.2.1,
the following conditions are satisfied:
(i) v1(t)f(0, t) 6= 0, t ∈ [0, T ];
(ii) µ1(0) = u0(0), µ2(0) = u′0(L), v1(0)u′0(0) + v2(0)u0(L) = k(0).
Then the inverse problem(2.1), (2.2), (2.7), (2.8)representing Case 2 is uniquely solv-
able.
Theorem 2.2.3 Assume that, in addition to conditions(2.4), (i) and (ii) of Theorem
2.2.1, the following conditions are satisfied:
µ1(0) = u0(0), µ2(0) = u0(L), v1(0)u′0(0) + v2(0)u
′0(L) = k(0).
Then the inverse problem(2.1), (2.2), (2.9), (2.10) representing Case 3 is uniquely
solvable.
Theorem 2.2.4 Assume that the regularity conditions(2.4)are satisfied and that:
(i) v1 ∈ C[0, T ], v2, µi ∈ C1[0, T ], i = 1, 2, v1(t) > 0, t ∈ [0, T ];
Chapter 2. 21
(ii) in Case 4,
f(0, t)− v2(t)f(L, t) 6= 0, t ∈ [0, T ],
µ1(0) = u0(0), µ2(0) = u′0(L) + v1(0)u0(L), k(0) = u′0(0) + v2(0)u′0(L);
(iii) in Case 5,
f(0, t) + v2(t)f(L, t) 6= 0, t ∈ [0, T ],
µ1(0) = u′0(0), µ2(0) = u′0(L) + v1(0)u0(L), k(0) = u0(0) + v2(0)u0(L).
Then the inverse problem(2.1), (2.2), (2.11), (2.12) representing Case 4, and(2.1),
(2.2), (2.13), (2.14)representing Case 5 are uniquely solvable.
Theorem 2.2.5 Assume that the regularity conditions(2.4)are satisfied and that:
(i) vi ∈ C[0, T ], v3, v4, µi ∈ C1[0, T ], vi(t) > 0, i = 1, 2, t ∈ [0, T ];
(ii) v1(t)f(0, t) + v2(t)f(L, t) 6= 0, v23(t) + v24(t) > 0, t ∈ [0, T ];
(iii) µ1(0) = u′0(0)− v1(0)u0(0), µ2(0) = u′0(L) + v2(0)u0(L),
k(0) = v3(0)u0(0) + v4(0)u0(L).
Then the inverse problem(2.1), (2.2), (2.15), (2.16) representing Case 6 is uniquely
solvable.
Although the problems of Cases 1–6 are uniquely solvable, they are still ill-posed
since small errors in the input datak(t) lead to large errors in the output source solution
r(t). In the next subsection we describe how the BEM discretisingnumerically the heat
equation (2.1) can be used in conjunction with the Tikhonov regularisation in order to
obtain a stable solution.
Chapter 2. 22
2.3 The boundary element method (BEM)
In the numerical process, we employ the BEM as introduced in Section 1.3. For the
heat equation (2.1) we then obtain the integral equation
η(x)u(x, t) =
∫ t
0
[
G(x, t, ξ, τ)∂u
∂n(ξ)(ξ, τ)− u(ξ, τ)
∂G
∂n(ξ)(x, t, ξ, τ)
]
ξ∈0,L
dτ
+
∫ L
0
G(x, t, y, 0)u(y, 0) dy+
∫ L
0
∫ t
0
G(x, t, y, τ)r(τ)f(y, τ) dτdy
+
∫ L
0
∫ t
0
G(x, t, y, τ)h(y, τ) dτdy, (x, t) ∈ [0, L]× (0, T ). (2.17)
We use the constant BEM (CBEM) with the midpoint approximations (1.12), (1.13)
and (1.15). Nevertheless, higher-order, e.g. linear boundary element approximations
will be more accurate than constant boundary elements. Thisimprovement in accuracy
will be significant in higher-dimension, see e.g. [49], but in our one-dimensional time-
dependent setting the use of the CBEM approximation was found sufficiently accurate.
With this, the integral equation (2.17) can be approximatedas
η(x)u(x, t) =
N∑
j=1
[A0j(x, t)q0j + ALj(x, t)qLj − B0j(x, t)h0j − BLj(x, t)hLj]
+
N0∑
k=1
Ck(x, t)u0,k + d(x, t) + d0(x, t), (2.18)
where the coefficientsAξj, Bξj, ξ ∈ 0, 1, andCk are given by (1.18)–(1.20) and can
be evaluated analytically as in (1.22)–(1.24), respectively. Whereas the double integral
source termsd andd0 are given by
d(x, t) =
∫ L
0
∫ t
0
G(x, t, y, τ)r(τ)f(y, τ) dτdy, (2.19)
d0(x, t) =
∫ L
0
∫ t
0
G(x, t, y, τ)h(y, τ) dτdy, (2.20)
Chapter 2. 23
and can be evaluated by assuming the piecewise constant approximations for the source
functionsf(x, t), h(x, t), andr(t), i.e.
f(x, t) = f(x, tj), h(x, t) = h(x, tj), r(t) = r(tj) =: rj, (2.21)
for t ∈ (tj−1, tj] and j = 1, N . By these approximations, the integrals (2.19) and
(2.20) become
d(x, t) =
∫ t
0
r(τ)
∫ L
0
G(x, t, y, τ)f(y, τ) dydτ =
N∑
j=1
Dj(x, t)rj, (2.22)
d0(x, t) =
∫ t
0
∫ L
0
G(x, t, y, τ)h(y, τ) dydτ =N∑
j=1
D0,j(x, t), (2.23)
where
Dj(x, t) =
∫ tj
tj−1
∫ L
0
G(x, t, y, τ)f(y, tj) dydτ =
∫ L
0
f(y, tj)
∫ tj
tj−1
G(x, t, y, τ) dτdy
=
∫ L
0
f(y, tj)Ayj(x, t) dy,
D0,j(x, t) =
∫ tj
tj−1
∫ L
0
G(x, t, y, τ)h(y, tj) dydτ =
∫ L
0
h(y, tj)
∫ tj
tj−1
G(x, t, y, τ) dτdy
=
∫ L
0
h(y, tj)Ayj(x, t) dy,
are evaluated numerically using the midpoint integral approximation. Here, the integral
equation (2.18) can be rewritten as
η(x)u(x, t) =
N∑
j=1
[A0j(x, t)q0j + ALj(x, t)qLj −B0j(x, t)h0j −BLj(x, t)hLj ]
+
N0∑
k=1
Ck(x, t)u0,k +
N∑
j=1
Dj(x, t)rj +
N∑
j=1
D0,j(x, t). (2.24)
Chapter 2. 24
Applying (2.24) at the boundary nodes(0, ti) and(L, ti) for i = 1, N , we obtain the
following system of2N equations
Aq¯−Bh
¯+ Cu
¯0+Dr
¯+ d
¯= 0
¯, (2.25)
where matricesA, B, C, D and vectors q¯, h
¯, u
¯0, r¯, d
¯are defined as same as in Section
1.4, andD =
Dj(0, ti)
Dj(L, ti)
2N×N
.
In this section, we consider the heat equation (2.1) with thegeneral conditions
(2.3) which can be separated into the 6 cases presented in (2.5)–(2.16). Applying the
boundary and the overdetermination conditions of these 6 cases results as follows.
2.3.1 Case 1
The Neumann heat flux boundary conditions (2.5) give
q¯=
−ux(0, tj)ux(L, tj)
2N
=
−µ1(tj)
µ2(tj)
2N
. (2.26)
Also, from (2.25) we obtain
h¯= B−1
(
Aq¯+ Cu
¯0+Dr
¯+ d
¯
)
. (2.27)
The overdetermination condition (2.6) can be rewritten as amatrix equation as follows:
[
V1 V2
]
h¯= k
¯, (2.28)
whereV1, V2 areN × N diagonal matrices of componentsv1(t1), . . . , v1(tN ) and
v2(t1), . . . , v2(tN), respectively, and k¯
is anN-column vector of the piecewise con-
stant approximation ofk(t), namely
k(t) = k(ti) =: ki, for t ∈ (tj−1, tj], j = 1, N.
Chapter 2. 25
Substituting (2.27) into (2.28) yields
[
V1 V2
]
B−1(
Aq¯+ Cu
¯0+Dr
¯+ d
¯
)
= k¯.
Rearranging this expression the inverse source problem in Case 1 reduces to solving
theN ×N linear system of equations
X1r¯= y
¯1, (2.29)
whereX1 =[
V1 V2
]
B−1D, and y¯1
= k¯−[
V1 V2
]
B−1(
Aq¯+ Cu
¯0+ d
¯
)
.
2.3.2 Case 2
We rearrange the matrix equation (2.25) as follows:
[
A0 AL
]
−ux(0, tj)ux(L, tj)
−[
B0 BL
]
u(0, tj)
u(L, tj)
+ Cu¯0
+Dr¯+ d
¯= 0
¯, (2.30)
whereA0 =
A0j(0, tj)
A0j(L, tj)
, AL =
ALj(0, tj)
ALj(L, tj)
, B0 =
B0j(0, tj) +12δij
B0j(L, tj)
,
andBL =
BLj(0, tj)
BLj(L, tj) +12δij
are2N × N matrices. Next, we apply the boundary
conditions (2.7) such that this system becomes
A0q¯0
+ ALµ¯2
−B0µ¯1
− BLh¯L
+ Cu¯0
+Dr¯+ d
¯= 0
¯,
whereµ¯1
=[
µ1(tj)]
N, µ
¯2=[
µ2(tj)]
N, q
¯0=[
q0j
]
N, and h
¯L=[
hLj
]
N. From this
system we obtain
−q¯0
h¯L
=[
A0 BL
]−1 (
ALµ¯2
− B0µ¯1
+ Cu¯0
+Dr¯+ d
¯
)
. (2.31)
Chapter 2. 26
The overdetermination condition (2.8) can be written in matrix equation as follows:
[
V1 V2
]
−q¯0
h¯L
= k¯. (2.32)
Substituting the expression (2.31) into (2.32) we obtain
[
V1 V2
] [
A0 BL
]−1 (
ALµ¯2
− B0µ¯1
+ Cu¯0
+Dr¯+ d
¯
)
= k¯.
Rearranging this expression the inverse source problem reduces to solving theN ×N
linear system of equations
X2r¯= y
¯2, (2.33)
where X2 =[
V1 V2
] [
A0 BL
]−1
D
and y¯2
= k¯−[
V1 V2
] [
A0 BL
]−1 (
ALµ¯2
− B0µ¯1
+ Cu¯0
+ d¯
)
.
2.3.3 Case 3
The Dirichlet boundary temperature conditions (2.9) give
h¯=
u(0, tj)
u(L, tj)
2N
=
µ1(tj)
µ2(tj)
2N
. (2.34)
Also, from (2.25) we obtain
q¯= A−1 (Bh
¯− Cu
¯0−Dr
¯− d
¯) . (2.35)
The overdetermination condition (2.10) can be written as
[
−V1 V2
]
q¯= k
¯. (2.36)
Chapter 2. 27
Substituting (2.35) into (2.36) gives theN ×N linear system of equations
X3r¯= y
¯3, (2.37)
whereX3 =[
−V1 V2
]
A−1D and y3 = −k¯+[
−V1 V2
]
A−1 (Bh¯− Cu
¯0− d
¯).
2.3.4 Case 4
Consider the boundary condition (2.11) which can be rewritten as
ux(L, t) = µ2(t)− v1(t)u(L, t).
Apply this to the matrix equation (2.30), derived from (2.25), then the system becomes
A0q¯+ AL(µ
¯2− V1h
¯L)− B0µ
¯1−BLh
¯L+ Cu
¯0+Dr
¯+ d
¯= 0
¯.
We rearrange the matrix equation above as follows:
[
A0 ALV1 +BL
]
−q¯0
h¯L
= ALµ¯2
− B0µ¯1
+ Cu¯0
+Dr¯+ d
¯.
From this system we obtain
−q¯0
h¯L
=[
A0 ALV1 +BL
]−1 (
ALµ¯2
− B0µ¯1
+ Cu¯0
+Dr¯+ d
¯
)
. (2.38)
The overdetermined condition (2.12) becomesux(0, t)+v2(t)µ2(t)−v2(t)v1(t)u(L, t) =k(t) and can be rewritten in the matrix form as,
[
I −V2V1]
−q¯0
h¯L
= k¯− V2µ
¯2, (2.39)
Chapter 2. 28
whereI is theN × N identity matrix. We then substitute (2.38) into (2.39) to obtain
theN ×N linear system of equations
X4r¯= y
¯4, (2.40)
where X4 =[
I −V2V1] [
A0 ALV1 +BL
]−1
D
and y¯4
= k¯−V2µ
¯2−[
I −V2V1] [
A0 ALV1 +BL
]−1 (
ALµ¯2
−B0µ¯1
+ Cu¯0
+ d¯
)
.
2.3.5 Case 5
Consider the boundary condition (2.13) which can be rewritten as
ux(L, t) = µ2(t)− v1(t)u(L, t). (2.41)
Apply this to the matrix equation (2.30), derived from (2.25), then the system becomes
−A0µ¯1
+ AL(µ¯2
− V1h¯L)− B0h
¯0−BLh
¯L+ Cu
¯0+Dr
¯+ d
¯= 0
¯.
We rearrange the matrix equation above as follows:
[
B0 ALV1 +BL
]
h¯= −A0µ
¯1+ ALµ
¯2+ Cu
¯0+Dr
¯+ d
¯.
From this system we obtain
h¯=[
B0 ALV1 +BL
]−1 (
−A0µ¯1
+ ALµ¯2
+ Cu¯0
+Dr¯+ d
¯
)
. (2.42)
The overdetermination condition (2.14) gives
[
I V2
]
h¯= k
¯. (2.43)
Chapter 2. 29
Substitute (2.42) into (2.43) to obtain theN ×N linear system of equations
X5r¯= y
¯5, (2.44)
where X5 =[
I V2
] [
B0 ALV1 +BL
]−1
D
and y¯5
= k¯−[
I V2
] [
B0 ALV1 +BL
]−1 (
−A0µ¯1
+ ALµ¯2
+ Cu¯0
+ d¯
)
.
2.3.6 Case 6
Consider the boundary condition (2.15) which can be rearranged as
q¯=
−ux(0, tj)ux(L, tj)
2N
=
−µ¯1
− V1u(0, tj)
µ¯2
− V2u(L, tj)
2N
.
Substituting this into the matrix equation (2.30) we obtain
A0(−µ¯1
− V1h¯0) + AL(µ
¯2− V2h
¯L)− B0h
¯0−BLh
¯L+ Cu
¯0+Dr
¯+ d
¯= 0
¯,
and this can be rearranged as
[
A0V1 +B0 ALV2 +BL
]
h¯= −A0µ
¯1+ ALµ
¯2+ Cu
¯0+Dr
¯+ d
¯.
Then we have
h¯=[
A0V1 +B0 ALV2 +BL
]−1 (
−A0µ¯1
+ ALµ¯2
+ Cu¯0
+Dr¯+ d
¯
)
. (2.45)
The overspecified condition (2.16) gives
[
V3 V4
]
h¯= k
¯. (2.46)
Chapter 2. 30
Substitute (2.45) into (2.46) to obtain theN ×N linear system of equations
X6r¯= y
¯6, (2.47)
where[
V3 V4
] [
A0V1 +B0 ALV2 +BL
]−1
D
and k¯−[
V3 V4
] [
A0V1 +B0 ALV2 +BL
]−1 (
−A0µ¯1
+ ALµ¯2
+ Cu¯0
+ d¯
)
.
From the above assembly one can see that the solution of the inverse heat source
problem (2.1)–(2.3) separated into the 6 cases, has been reduced to solving theN ×N
linear system of equations, generally written as
Xr¯= y
¯, (2.48)
whereX and y¯
are the coefficient matrix and the right-hand side vector, respectively,
corresponding to the case we are dealing with; that is, the linear systems of equations,
(2.29), (2.33), (2.37), (2.40), (2.44), and (2.47) for Cases 1–6, respectively. We note
that the system of equations (2.48) is ill-conditioned since the inverse problems un-
der investigation are ill-posed. Therefore, a straightforward inversion of (2.48) such
as the Gaussian elimination or the singular value decomposition will result into an
unstable numerical solution, especially when the right-hand side vector y¯
is contami-
nated by random noise as y¯ǫ = y
¯+ ǫ whereǫ represents the noise to contaminate into
the problem. In order to ensure a stable solution we employ the Tikhonov regulari-
sation method for (2.48) which gives solution (1.34) together with the regularisation
parameter chosen by the discrepancy principle. Let us denote byλdis the regularisa-
tion parameter which is determined by the discrepancy principle, i.e. the largestλ for
which the residual‖Xr¯− yǫ‖ becomes less than the noise levelǫ.
Chapter 2. 31
2.4 Numerical examples and discussion
In order to test the accuracy of the approximations, let us introduce the root mean
square error (RMSE) defined as
RMSE(r(t)) =
√
√
√
√
T
N
N∑
i=1
(
rexact(ti)− rnumerical(ti))2. (2.49)
2.4.1 Example 1
We consider a smooth benchmark test with the input data
u(x, 0) = u0(x) = 1 + x− x2,
f(x, t) = (1− x2)e−t, h(x, t) = (2 + x)et.
(2.50)
Assuming that all quantities involved have been non-dimensionalised we can takeT =
L = 1. In addition, the boundary and overdetermination conditions are as follows:
Case 1 ux(0, t) = et, ux(1, t) = −et, (2.51)
u(0, t) + u(1, t) = 2et. (2.52)
Case 2 u(0, t) = et, ux(1, t) = −et, (2.53)
ux(0, t) + u(1, t) = 2et. (2.54)
Case 3 u(0, t) = et, u(1, t) = et, (2.55)
etux(0, t) + tux(1, t) = (et − t)et, (2.56)
where v1(t) = et, v2(t) = t.
Case 4 u(0, t) = et, ux(1, t) + (1 + t)u(1, t) = tet, (2.57)
ux(0, t) + e−tux(1, t) = et − 1, (2.58)
where v1(t) = 1 + t, v2(t) = e−t.
Chapter 2. 32
Case 5 ux(0, t) = et, ux(1, t) + e−tu(1, t) = 1− et, (2.59)
u(0, t) + (1 + t)u(1, t) = (2 + t)et, (2.60)
where v1(t) = e−t, v2(t) = 1 + t.
Case 6 ux(0, t)− etu(0, t) = et − e2t, ux(1, t) + (1 + t)u(1, t) = tet, (2.61)
tu(0, t) + (1− t)u(1, t) = et, (2.62)
where v1(t) = et, v2(t) = 1 + t, v3(t) = t, v4(t) = 1− t.
In this example the analytical solution is given by
u(x, t) = (1 + x− x2)et, r(t) = e2t, (2.63)
for 0 ≤ x ≤ 1 and0 ≤ t ≤ 1. Note that the input data in expressions (2.50)–(2.62)
satisfy the conditions of Theorems 2.2.1–2.2.5 for the existence and uniqueness of the
solution of the inverse problems of Cases 1–6 under investigation.
Figure 2.1 and Table 2.1 show the condition numbers of matrixX in (2.48) cor-
responding toN0 = N ∈ 20, 40, 80 for all Cases 1–6. From these it can be seen
that the condition numbers of the matrixX for Cases 1, 5, and, 6 are high and this
ill-conditioning will need to be dealt with using the Tikhonov regularisation described
in Section 1.6.2. On the other hand, the matrices for Cases 2–4 do not have a very large
condition number and, in principle, the system of equations(2.48) can be solved di-
rectly using, for example, the Gauss elimination method or the SVD. In what follows,
we illustrate the numerical results obtained withN0 = N = 40.
Case 1
Figures 2.2(a)–2.2(c) show the analytical and numerical solutions forr(t), u(0, t) and
u(1, t), respectively, for exact input data and no regularisation,i.e. λ = 0. From
this figure it can be seen that although the solution for the boundary values of the
temperaturesu(0, t) andu(1, t) is stable and accurate, the retrieved source termr(t)
Chapter 2. 33
0 10 20 30 40 50 60 70 8010
−7
10−6
10−5
10−4
10−3
10−2
10−1
100
Cond = 6.19E+3
Nor
mal
ised
Sin
gula
r V
alue
s
N
Cond = 1.26E+5
Cond = 7.49E+6
(a)
0 10 20 30 40 50 60 70 8010
−3
10−2
10−1
100
Cond = 87
Nor
mal
ised
Sin
gula
r V
alue
s
N
Cond = 248
Cond = 705
(b)
0 10 20 30 40 50 60 70 8010
−3
10−2
10−1
100
Cond = 22
Nor
mal
ised
Sin
gula
r V
alue
s
N
Cond = 54
Cond = 125
(c)
0 10 20 30 40 50 60 70 8010
−3
10−2
10−1
100
Cond = 38
Nor
mal
ised
Sin
gula
r V
alue
s
N
Cond = 107
Cond = 298
(d)
0 10 20 30 40 50 60 70 8010
−10
10−8
10−6
10−4
10−2
100
Cond = 1.52E+4
Nor
mal
ised
Sin
gula
r V
alue
s
N
Cond = 7.36E+5
Cond = 1.76E+8
(e)
0 10 20 30 40 50 60 70 8010
−10
10−8
10−6
10−4
10−2
100
Cond = 1.61E+4
Nor
mal
ised
Sin
gula
r V
alue
s
N
Cond = 2.52E+6
Cond = 5.89E+9
(f)
Figure 2.1: The normalised singular values of matrixX for (a) Case 1 – (f) Case 6 forN0 = N = 20 (− · −), 40 (· · · ), 80 (−−−), for Example 1.
Chapter 2. 34
Table 2.1: The condition numbers of the matrixX in equation (2.48) forN = N0 ∈20, 40, 80 for Example 1 Cases 1–6.
CaseCondition Number
N = N0 = 20 N = N0 = 40 N = N0 = 801 6.19E+3 1.26E+5 7.49E+62 87 248 7053 22 54 1254 38 107 2985 1.52E+4 7.36E+5 1.76E+86 1.61E+4 2.52E+6 5.89E+9
seems unstable. This is to be expected since the inverse problem in Case 1 is ill-
posed, see also the condition number in Table 2.1. Regularisation needs to be employed
and more stable results are illustrated in Figure 2.3. In this figure, numerical results
obtained with the zeroth-order Tikhonov regularisation. Initially, we have tried the
L-curve criterion, but we have found that anL-corner could not be clearly identified.
We then have tried with the trial and error and found that the regularisation parameters
λ ∈ 10−7, 10−5 are most suitable as presented in Figure 2.3. Moreover, although
not illustrated, it is reported that the slight inaccuracies neart = 1 observed in Figure
2.3(a) can be further eliminated by employing higher-orderregularisations such as the
first- or second-order.
Table 2.2: The RMSE for the zeroth- and first-order Tikhonov regularisation forp ∈0, 1, 3% noise, for Example 1 Case 1.
Regularisation p(%)parameter RMSE
λ r(t) u(0, t) u(1, t)- 0 0 0.983 3.67E-4 3.67E-4- 1 0 4.90E+2 1.90E-1 1.81E-1
zeroth-order0 λ=1.0E-5 0.499 1.27E-3 7.92E-51 λdis=6.6E-4 1.264 2.01E-2 7.31E-33 λdis=4.0E-3 1.982 6.25E-2 2.91E-2
first-order0 λ=1.0E-7 9.49E-3 3.06E-5 3.07E-51 λdis=4.3E-2 0.364 7.19E-3 3.03E-33 λdis=9.7E-1 1.007 3.59E-2 1.78E-2
In order to investigate the stability of the numerical solution we add noise to the
Chapter 2. 35
0 0.2 0.4 0.6 0.8 11
2
3
4
5
6
7
8
9
10
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(0
,t)
(b)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(1
,t)
(c)
Figure 2.2: The analytical (—–) and numerical (− · −) results of (a)r(t), (b) u(0, t),and (c)u(1, t) for exact data andλ = 0, for Example 1 Case 1.
right-hand side of the overspecified condition (2.6) as
k¯ǫ = k
¯+ random(′Normal′, 0, σ, 1, N), (2.64)
where therandom(′Normal′, 0, σ, 1, N) is a command in MATLAB which generates
the random variables by normal distribution with zero mean and standard deviationσ,
computed as
σ = p× maxt∈[0,1]
|k(t)|, (2.65)
wherep represents the percentage of noise. In Case 1,k is given by (2.52) and there-
fore,σ = 2ep in (2.65). The ill-posedness of the inverse problem and the instability of
Chapter 2. 36
0 0.2 0.4 0.6 0.8 11
2
3
4
5
6
7
8
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 11
2
3
4
5
6
7
8
t
r(t
)
(b)
Figure 2.3: The analytical (—–) and numerical (− · −) results ofr(t) obtained byusing the zeroth-order Tikhonov regularisation with the regularisation parameters (a)λ = 10−7 gives RMSE=0.319, and (b)λ = 10−5 gives RMSE=0.499, for exact datafor Example 1 Case 1.
the numerical solution in Case 1 are further enhanced by the presence of noise in the
measured data, as illustrated in Figure 2.4 (compare with Figure 2.2). In order to re-
move the highly unwanted oscillations recorded in Figure 2.4, we employ the Tikhonov
regularisation with the choice of the regularisation parameterλ, as described in (1.34),
given by the discrepancy principle as introduced in Section1.6.3. Figures 2.5(a) and
2.6(a) present the discrepancy principle curves obtained by the Tikhonov regularisa-
tion of order zero and one, respectively, forp = 1% and3% noisy data. Generated
as in (2.64), this results in the amounts of noiseǫ = 0.32 and 1.01 forp = 1% and
3%, respectively. The intersections between these horizontal lines and the discrepancy
(residual) curves yield the regularisation parameter denoted byλdis and tabulated in
Table 2.2. With these values ofλdis, Figures 2.5(b)–2.5(d) and 2.6(b)–2.6(d) present
the numerical results forr(t), u(0, t), andu(1, t) obtained using the zeroth- and first-
order Tikhonov regularisation, respectively. By comparing Figures 2.5(b) and 2.6(b)
it can be seen that the first-order regularisation produces more accurate and stable re-
sults than the zeroth-order regularisation since it imposes a higher-order smoothness
constraint onto the solution.
Chapter 2. 37
0 0.2 0.4 0.6 0.8 1−1500
−1000
−500
0
500
1000
1500
2000
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 11
1.5
2
2.5
3
3.5
t
u(0
,t)
(b)
0 0.2 0.4 0.6 0.8 11
1.5
2
2.5
3
3.5
t
u(1
,t)
(c)
Figure 2.4: The analytical (—–) and numerical (− · −) results of (a)r(t), (b) u(0, t),and (c)u(1, t) for p = 1% noisy data andλ = 0, for Example 1 Case 1.
Cases 2–4
In Cases 2–4, Table 2.1 shows that the condition numbers of the matricesX2, X3, and
X4 are much smaller than the condition numbers of the matrices for the rest of the
cases. Therefore, for exact data, i.e.p = 0, we expect accurate and stable results
of the inversion even if no regularisation is employed, i.e.λ = 0. This is clearly
illustrated in Figure 2.7 for Case 2 where it can be seen that the agreement between
the analytical and numerical solutions forr(t), u(1, t) andux(0, t) is excellent. The
same overlapping agreement has also been obtained for Cases3 and 4 and therefore
these results are omitted. Next we addp% noise in the input data, as described in
Chapter 2. 38
10−6
10−5
10−4
10−3
10−2
10−1
100
10−2
10−1
100
101
ε = 0.32λ = 6.6E−4
ε = 1.01λ = 4.0E−3
λ
‖X
r ¯λ−
y ¯ǫ‖
(a)
0 0.2 0.4 0.6 0.8 10
1
2
3
4
5
6
7
8
t
r(t
)
(b)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(0
,t)
(c)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(1
,t)
(d)
Figure 2.5: (a) The discrepancy principle curve, and the analytical (—–) and numericalresults of (b)r(t), (c)u(0, t), and (d)u(1, t) obtained using the zeroth-order Tikhonovregularisation forp = 1% (− · −) andp = 3% (− − −) noise with the regularisationparametersλdis given in Table 2.2, for Example 1 Case 1.
(2.64). According to expression (2.65), and equations (2.54), (2.56), and (2.58), we
have the standard deviationsσ = 2ep, σ = (e2 − e)p, andσ = (e − 1)p for Cases
2–4, respectively. As expected, and previously reported inFigure 2.4 for Case 1, if
no regularisation is imposed, i.e.λ = 0, whenp = 1% noise contaminates the input
datak(t), then an unstable solution forr(t) is obtained, see Figures 2.8, 2.11, and
2.14 for Cases 2–4, respectively. However, the high oscillations in these figures have
smaller magnitudes than these reported in Figure 2.4 for Case 1. This is consistent
with the fact that the inverse problems of Cases 2–4 are less ill-posed than the inverse
Chapter 2. 39
10−4
10−3
10−2
10−1
100
101
102
103
0.5
1
1.5
2
ε = 0.32λ = 4.3E−2
ε = 1.01λ = 9.7E−1
λ
‖X
r ¯λ−
y ¯ǫ‖
(a)
0 0.2 0.4 0.6 0.8 10
1
2
3
4
5
6
7
8
t
r(t
)
(b)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(0
,t)
(c)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(1
,t)
(d)
Figure 2.6: (a) The discrepancy principle curve, and the analytical (—–) and numericalresults of (b)r(t), (c) u(0, t), and (d)u(1, t) obtained using the first-order Tikhonovregularisation forp = 1% (− · −) andp = 3% (− − −) noise with the regularisationparametersλdis given in Table 2.2, for Example 1 Case 1.
problem of Case 1 (and Cases 5 and 6), see the condition numbers reported in Table
2.1. It is also interesting to remark that the retrieval of the boundary temperature
is more accurate and stable than of the heat flux, e.g. compareFigures 2.8(b) and
2.14(b) with Figures 2.8(c) and 2.14(c), see also Figures 2.11(b) and 2.11(c). This is
to be expected since retrieving higher-order derivatives is less accurate than retrieving
lower-order derivatives, see Lesnic et al. [38]. In order toobtain a stable solution,
regularisation needs to be employed. The numerical resultsobtained forp ∈ 1, 3%noise using the zeroth-order Tikhonov regularisation withthe regularisation parameter
Chapter 2. 40
chosen according to the discrepancy principle are shown in Figures 2.9, 2.12, and
2.15 for Cases 2–4, respectively. The corresponding results obtained using the first-
order regularisation illustrated in Figures 2.10, 2.13, and 2.16 show much improvement
in terms of both stability and accuracy compared to the results obtained using the
zeroth-order regularisation. The regularisation parameters and the RMSE errors of the
output solutions are given in Tables 2.3–2.5 for Cases 2–4, respectively. From these
tables it can be observed that, as expected,λdis and RMSE increase with increasing the
percentage of noisep.
0 0.2 0.4 0.6 0.8 11
2
3
4
5
6
7
8
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(1
,t)
(b)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
ux(0
,t)
(c)
Figure 2.7: The analytical (—–) and numerical (− · −) results of (a)r(t), (b) u(1, t),and (c)ux(0, t) for exact data andλ = 0, for Example 1 Case 2.
Chapter 2. 41
0 0.2 0.4 0.6 0.8 1−6
−4
−2
0
2
4
6
8
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(1
,t)
(b)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
ux(0
,t)
(c)
Figure 2.8: The analytical (—–) and numerical (− · −) results of (a)r(t), (b) u(1, t),and (c)ux(0, t) for p = 1% noisy data andλ = 0, for Example 1 Case 2.
Cases 5 and 6
In Cases 5 and 6 we expect even higher ill-conditioning to occur in the systems of equa-
tions (2.44) and (2.47), compare the condition numbers in Table 2.1. This is reflected
indeed in the numerical results presented in Figures 2.17(a) and 2.18 for Case 5, and
even more prominently in Figure 2.21 for Case 6 where unstable numerical results can
be clearly seen ifλ = 0 even when exact data are inverted. Whenp% noise is added
to the measured data (2.60) and (2.62), the standard deviations in (2.65) are given by
σ = 3ep andσ = ep for Cases 5 and 6, respectively. Numerical results obtainedusing
the zeroth- and first-order regularisation are presented inFigures 2.19, 2.20 and Table
Chapter 2. 42
10−6
10−5
10−4
10−3
10−2
10−1
100
101
10−3
10−2
10−1
100
101
ε = 0.34λ = 2.8E−3
ε = 1.02λ = 8.8E−3
λ
‖X
r ¯λ−
y ¯ǫ‖
(a)
0 0.2 0.4 0.6 0.8 10
1
2
3
4
5
6
7
8
t
r(t
)
(b)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(1
,t)
(c)
0 0.2 0.4 0.6 0.8 10.8
1
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
ux(0
,t)
(d)
Figure 2.9: (a) The discrepancy principle curve, and the analytical (—–) and numericalresults of (b)r(t), (c)u(1, t), and (d)ux(0, t) obtained using the zeroth-order Tikhonovregularisation forp = 1% (− · −) andp = 3% (− − −) noise with the regularisationparametersλdis given in Table 2.3, for Example 1 Case 2.
Table 2.3: The RMSE for the zeroth- and first-order Tikhonov regularisation forp ∈0, 1, 3% noise, for Example 1 Case 2.
Regularisation p(%)parameter RMSEλdis r(t) u(1, t) ux(0, t)
- 0 0 7.15E-3 4.33E-5 4.33E-5- 1 0 3.29 5.94E-3 5.31E-2
zeroth-order1 2.8E-3 0.785 7.63E-3 5.83E-23 8.8E-3 1.331 1.66E-2 1.24E-1
first-order1 3.0E-1 3.23E-1 3.29E-3 2.95E-23 9.4E-1 5.53E-1 8.58E-3 5.81E-2
Chapter 2. 43
10−4
10−3
10−2
10−1
100
101
102
103
104
105
100
ε = 0.34λ = 3.0E−1
ε = 1.02λ = 9.4E−1
λ
‖X
r ¯λ−
y ¯ǫ‖
(a)
0 0.2 0.4 0.6 0.8 10
1
2
3
4
5
6
7
8
t
r(t
)
(b)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(1
,t)
(c)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
ux(0
,t)
(d)
Figure 2.10: (a) The discrepancy principle curve, and the analytical (—–) and nu-merical results of (b)r(t), (c) u(1, t), and (d)ux(0, t) obtained using the first-orderTikhonov regularisation forp = 1% (− · −) andp = 3% (− − −) noise with theregularisation parametersλdis given in Table 2.3, for Example 1 Case 2.
Chapter 2. 44
0 0.2 0.4 0.6 0.8 1−4
−2
0
2
4
6
8
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
ux(0
,t)
(b)
0 0.2 0.4 0.6 0.8 1−2.8
−2.6
−2.4
−2.2
−2
−1.8
−1.6
−1.4
−1.2
−1
t
ux(1
,t)
(c)
Figure 2.11: The analytical (—–) and numerical (−·−) results of (a)r(t), (b)ux(0, t),and (c)ux(1, t) for p = 1% noisy data andλ = 0, for Example 1 Case 3.
2.6 for Case 5, whilst for Case 6 the corresponding results are presented in Figures
2.22, 2.23 and Table 2.7. Furthermore, for Case 6, which is the most ill-conditioned
case, we have increased the percentage of noise top = 5% in order to show that the
Tikhonov regularisation method combined with the BEM can satisfactorily deal in a
stable and accurate manner with higher measurement errors.We finally report that,
although not illustrated, we have also implemented the second-order Tikhonov reg-
ularisation and similar results, in terms of accuracy and stability, to those given by
first-order regularisation have been obtained.
Chapter 2. 45
0 0.0001 0.001 0.01 0.1 1 10 100 1000 1000010
−3
10−2
10−1
100
101
ε = 0.29λ = 3.2E−3
ε = 0.84λ = 8.7E−3
λ
‖X
r ¯λ−
y ¯ǫ‖
(a)
0 0.2 0.4 0.6 0.8 10
1
2
3
4
5
6
7
8
t
r(t
)
(b)
0 0.2 0.4 0.6 0.8 10.8
1
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
ux(0
,t)
(c)
0 0.2 0.4 0.6 0.8 1−2.8
−2.6
−2.4
−2.2
−2
−1.8
−1.6
−1.4
−1.2
−1
t
ux(1
,t)
(d)
Figure 2.12: (a) The discrepancy principle curve, and the analytical (—–) and numer-ical results of (b)r(t), (c) ux(0, t), and (d)ux(1, t) obtained using the zeroth-orderTikhonov regularisation forp = 1% (− · −) andp = 3% (− − −) noise with theregularisation parametersλdis given in Table 2.4, for Example 1 Case 3.
2.4.2 Example 2
We finally investigate specially in the most ill-posed Case 6, the retrieval of a non-
smooth heat source given by
r(t) =
∣
∣
∣
∣
t− 1
2
∣
∣
∣
∣
+ 1, 0 < t < 1 = T, (2.66)
Chapter 2. 46
10−4
10−3
10−2
10−1
100
101
102
103
104
105
10−2
10−1
100
101
ε = 0.29λ = 3.0E−1
ε = 0.84λ = 5.8E−1
λ
‖X
r ¯λ−
y ¯ǫ‖
(a)
0 0.2 0.4 0.6 0.8 11
2
3
4
5
6
7
8
t
r(t
)
(b)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
ux(0
,t)
(c)
0 0.2 0.4 0.6 0.8 1−2.8
−2.6
−2.4
−2.2
−2
−1.8
−1.6
−1.4
−1.2
t
ux(1
,t)
(d)
Figure 2.13: (a) The discrepancy principle curve, and the analytical (—–) and nu-merical results of (b)r(t), (c) ux(0, t), and (d)ux(1, t) obtained using the first-orderTikhonov regularisation forp = 1% (− · −) andp = 3% (− − −) noise with theregularisation parametersλdis given in Table 2.4, for Example 1 Case 3.
and, for brevity, the boundary and overdetermination conditions as given by (2.61),
(2.62) and the input data as follows:
u(x, 0) = u0(x) = 1 + x− x2, f(x, t) = (1− x2)e−t,
h(x, t) = (3 + x− x2)et − (1− x2)
(∣
∣
∣
∣
t− 1
2
∣
∣
∣
∣
+ 1
)
e−t,(2.67)
which are generated from the analytical temperature solution
u(x, t) = (1 + x− x2)et. (2.68)
Chapter 2. 47
Table 2.4: The RMSE for the zeroth- and first-order Tikhonov regularisation forp ∈0, 1, 3% noise, for Example 1 Case 3.
Regularisation p(%)parameter RMSEλdis r(t) ux(0, t) ux(1, t)
- 0 0 3.29E-3 4.04E-5 1.49E-4- 1 0 1.776 3.58E-2 2.41E-2
zeroth-order1 3.2E-3 4.42E-1 3.65E-2 1.74E-23 8.7E-3 7.80E-1 6.81E-2 3.09E-2
first-order1 3.0E-1 1.78E-1 1.72E-2 7.58E-33 5.8E-1 2.47E-1 2.41E-2 1.07E-2
0 0.2 0.4 0.6 0.8 11
2
3
4
5
6
7
8
9
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(1
,t)
(b)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
ux(0
,t)
(c)
0 0.2 0.4 0.6 0.8 1−2.8
−2.6
−2.4
−2.2
−2
−1.8
−1.6
−1.4
−1.2
−1
t
ux(1
,t)
(d)
Figure 2.14: The analytical (—–) and numerical (− · −) results of (a)r(t), (b)u(1, t),(c) ux(0, t), andux(1, t) for p = 1% noisy data andλ = 0, for Example 1 Case 4.
Numerical results are presented forN = N0 = 80. Also, results are illustrated for the
first-order Tikhonov regularisation only, as for the zeroth-order regularisation the re-
Chapter 2. 48
Table 2.5: The RMSE for the zeroth- and first-order Tikhonov regularisation forp ∈0, 1, 3% noise, for Example 1 Case 4.
Regularisation p(%)parameter RMSEλdis r(t) u(1, t) ux(0, t) ux(1, t)
- 0 0 4.13E-3 3.63E-5 3.19E-5 5.77E-5- 1 0 4.98E-1 9.11E-4 1.71E-2 1.33E-3
zeroth-order1 6.2E-4 4.35E-1 1.52E-3 1.89E-2 2.70E-33 2.0E-3 7.31E-1 4.55E-3 4.95E-2 8.24E-3
first-order1 4.8E-2 1.23E-1 9.29E-4 9.61E-3 1.38E-33 1.3E-1 1.77E-1 2.16E-3 1.72E-2 3.15E-3
Table 2.6: The RMSE for the zeroth- and first-order Tikhonov regularisation forp ∈0, 1, 3% noise, for Example 1 Case 5.
Regularisation p(%)parameter RMSEλdis r(t) u(0, t) u(1, t) ux(1, t)
- 0 0 5.598 3.35E-3 1.72E-3 6.76E-4- 1 0 2.9E+3 1.776 9.09E-1 3.56E-1
zeroth-order1 2.2E-3 1.584 3.92E-2 1.60E-2 7.79E-33 9.5E-3 2.033 9.38E-2 4.73E-2 2.53E-2
first-order1 1.7E-1 0.673 2.29E-2 1.02E-2 4.93E-33 3.4 1.334 7.52E-2 3.92E-2 2.00E-2
Table 2.7: The RMSE for the zeroth- and first-order Tikhonov regularisation forp ∈0, 1, 3, 5% noise, for Example 1 Case 6.
Regularisation p(%)parameter RMSEλdis r(t) u(0, t) u(1, t) ux(0, t) ux(1, t)
- 0 0 1.49E+2 2.58E-2 6.08E-2 4.66E-2 1.09E-1- 1 0 6.78E+4 1.17E+1 2.77E+1 2.13E+1 4.98E+1
zeroth-order1 1.8E-4 1.380 1.96E-2 9.67E-3 4.59E-2 1.68E-23 6.3E-4 1.559 4.03E-2 2.61E-2 8.14E-2 4.28E-25 8.8E-4 1.737 4.07E-2 2.23E-2 8.51E-2 3.78E-2
first-order1 2.4E-2 5.90E-1 1.19E-2 6.82E-3 2.59E-2 1.10E-23 8.4E-2 5.36E-1 1.71E-2 1.12E-2 2.81E-2 1.52E-25 9.5E-2 5.04E-1 1.80E-2 1.36E-2 3.49E-2 2.31E-2
Chapter 2. 49
10−6
10−5
10−4
10−3
10−2
10−1
100
101
102
103
10−4
10−3
10−2
10−1
100
101
ε = 0.10λ = 6.1E−4
ε = 0.31λ = 2.0E−3
λ
‖X
r ¯λ−
y ¯ǫ‖
(a)
0 0.2 0.4 0.6 0.8 10
1
2
3
4
5
6
7
8
t
r(t
)
(b)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(1
,t)
(c)
0 0.2 0.4 0.6 0.8 10.8
1
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
ux(0
,t)
(d)
0 0.2 0.4 0.6 0.8 1−2.8
−2.6
−2.4
−2.2
−2
−1.8
−1.6
−1.4
−1.2
t
ux(1
,t)
(e)
Figure 2.15: (a) The discrepancy principle curve, and the analytical (—–) and nu-merical results of (b)r(t), (c) u(1, t), (d) ux(0, t), and (e)ux(1, t) obtained using thezeroth-order Tikhonov regularisation forp = 1% (− · −) andp = 3% (− − −) noisewith the regularisation parametersλdis given in Table 2.5, for Example 1 Case 4.
Chapter 2. 50
0.0001 0.001 0.01 0.1 1 10 100 1000 1000010
−3
10−2
10−1
100
101
ε = 0.10λ = 4.1E−2
ε = 0.31λ = 1.3E−1
λ
‖X
r ¯λ−
y ¯ǫ‖
(a)
0 0.2 0.4 0.6 0.8 11
2
3
4
5
6
7
8
t
r(t
)
(b)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(1
,t)
(c)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
ux(0
,t)
(d)
0 0.2 0.4 0.6 0.8 1−2.8
−2.6
−2.4
−2.2
−2
−1.8
−1.6
−1.4
−1.2
t
ux(1
,t)
(e)
Figure 2.16: (a) The discrepancy principle curve, and the analytical (—–) and nu-merical results of (b)r(t), (c) u(1, t), (d) ux(0, t), and (e)ux(1, t) obtained using thefirst-order Tikhonov regularisation forp = 1% (− · −) andp = 3% (− − −) noisewith the regularisation parametersλdis given in Table 2.5, for Example 1 Case 4.
Chapter 2. 51
0 0.2 0.4 0.6 0.8 1−15
−10
−5
0
5
10
15
20
25
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(0
,t)
(b)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(1
,t)
(c)
0 0.2 0.4 0.6 0.8 1−2.8
−2.6
−2.4
−2.2
−2
−1.8
−1.6
−1.4
−1.2
−1
t
ux(1
,t)
(d)
Figure 2.17: The analytical (—–) and numerical (− · −) results of (a)r(t), (b)u(0, t),(c) u(1, t) and (d)ux(1, t) for exact data andλ = 0, for Example 1 Case 5.
Chapter 2. 52
0 0.2 0.4 0.6 0.8 1−1.5
−1
−0.5
0
0.5
1x 10
4
t
r(t
)
x 104
(a)
0 0.2 0.4 0.6 0.8 1−5
0
5
10
t
u(0
,t)
(b)
0 0.2 0.4 0.6 0.8 1−1
0
1
2
3
4
5
6
7
t
u(1
,t)
(c)
0 0.2 0.4 0.6 0.8 1−4
−3.5
−3
−2.5
−2
−1.5
−1
t
ux(1
,t)
(d)
Figure 2.18: The analytical (—–) and numerical (− · −) results of (a)r(t), (b) u(0, t),(c) u(1, t) and (d)ux(1, t) for p = 1% noisy data andλ = 0, for Example 1 Case 5.
Chapter 2. 53
0 0.0001 0.001 0.01 0.1 1 10 100 1000 10000
100
101
ε = 0.50λ = 2.2E−3
ε = 1.52λ = 9.5E−3
λ
‖X
r ¯λ−
y ¯ǫ‖
(a)
0 0.2 0.4 0.6 0.8 10
1
2
3
4
5
6
7
8
t
r(t
)
(b)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(0
,t)
(c)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(1
,t)
(d)
0 0.2 0.4 0.6 0.8 1−2.8
−2.6
−2.4
−2.2
−2
−1.8
−1.6
−1.4
−1.2
t
ux(1
,t)
(e)
Figure 2.19: (a) The discrepancy principle curve, and the analytical (—–) and nu-merical results of (b)r(t), (c) u(0, t), (d) u(1, t), and (e)ux(1, t) obtained using thezeroth-order Tikhonov regularisation forp = 1% (− · −) andp = 3% (− − −) noisewith the regularisation parametersλdis given in Table 2.6, for Example 1 Case 5.
Chapter 2. 54
10−3
10−2
10−1
100
101
102
103
104
105
106
11
2
ε = 0.50λ = 2.7E−1
ε = 1.52
λ = 3.4
λ
‖X
r ¯λ−
y ¯ǫ‖
(a)
0 0.2 0.4 0.6 0.8 11
2
3
4
5
6
7
8
t
r(t
)
(b)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(0
,t)
(c)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(1
,t)
(d)
0 0.2 0.4 0.6 0.8 1−2.8
−2.6
−2.4
−2.2
−2
−1.8
−1.6
−1.4
−1.2
−1
t
ux(1
,t)
(e)
Figure 2.20: (a) The discrepancy principle curve, and the analytical (—–) and nu-merical results of (b)r(t), (c) u(0, t), (d) u(1, t), and (e)ux(1, t) obtained using thefirst-order Tikhonov regularisation forp = 1% (− · −) andp = 3% (− − −) noisewith the regularisation parametersλdis given in Table 2.6, for Example 1 Case 5.
Chapter 2. 55
0 0.2 0.4 0.6 0.8 1−300
−200
−100
0
100
200
300
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(0
,t)
(b)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(1
,t)
(c)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
ux(0
,t)
(d)
0 0.2 0.4 0.6 0.8 1−3
−2.8
−2.6
−2.4
−2.2
−2
−1.8
−1.6
−1.4
−1.2
−1
t
ux(1
,t)
(e)
Figure 2.21: The analytical (—–) and numerical (− · −) results of (a)r(t), (b)u(0, t),(c) u(1, t), (d) ux(0, t), and (e)ux(1, t) for exact data andλ = 0, for Example 1 Case6.
Chapter 2. 56
10−6
10−5
10−4
10−3
10−2
10−1
100
101
102
103
10−2
10−1
100
101
ε = 0.17λ = 1.8E−4
ε = 0.47λ = 6.3E−4
ε = 0.85λ = 8.8E−4
λ
‖X
r ¯λ−
y ¯ǫ‖
(a)
0 0.2 0.4 0.6 0.8 1−1
0
1
2
3
4
5
6
7
8
t
r(t
)
(b)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(0
,t)
(c)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(1
,t)
(d)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
ux(0
,t)
(e)
0 0.2 0.4 0.6 0.8 1−2.8
−2.6
−2.4
−2.2
−2
−1.8
−1.6
−1.4
−1.2
−1
t
ux(1
,t)
(f)
Figure 2.22: (a) The discrepancy principle curve, and the analytical (—–) and numeri-cal results of (b)r(t), (c)u(0, t), (d)u(1, t), (e)ux(0, t), and (f)ux(1, t) obtained usingthe zeroth-order Tikhonov regularisation forp ∈ 1(−·−), 3(· · · ), 5(−−−)% noisewith the regularisation parametersλdis given in Table 2.7, for Example 1 Case 6.
Chapter 2. 57
10−4
10−3
10−2
10−1
100
101
102
103
104
105
10−0.8
10−0.6
10−0.4
10−0.2
100
ε = 0.17λ = 2.4E−2
ε = 0.47λ = 8.4E−2
ε = 0.85λ = 9.5E−2
λ
‖X
r ¯λ−
y ¯ǫ‖
(a)
0 0.2 0.4 0.6 0.8 10
1
2
3
4
5
6
7
8
t
r(t
)
(b)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(0
,t)
(c)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(1
,t)
(d)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
ux(0
,t)
(e)
0 0.2 0.4 0.6 0.8 1−2.8
−2.6
−2.4
−2.2
−2
−1.8
−1.6
−1.4
−1.2
−1
t
ux(1
,t)
(f)
Figure 2.23: (a) The discrepancy principle curve, and the analytical (—–) and numeri-cal results of (b)r(t), (c)u(0, t), (d)u(1, t), (e)ux(0, t), and (f)ux(1, t) obtained usingthe first-order Tikhonov regularisation forp ∈ 1(− · −), 3(· · · ), 5(−−−)% noisewith the regularisation parametersλdis given in Table 2.7, for Example 1 Case 6.
Chapter 2. 58
sults obtained were oscillatory as in Figure 2.22(b). Figure 2.24 shows the numerical
results obtained for various amounts of noisep ∈ 0, 0.1, 0.5, 1% for the regular-
isation parametersλdis given in Table 2.8. From this figure it can be seen that the
regularised BEM can invert accurately up to about0.1% noisy data. For higher levels
of noise the reconstruction of the non-smooth heat source (2.66) starts to deteriorate.
Table 2.8: The RMSE for the first-order Tikhonov regularisation for p ∈0, 0.1, 0.5, 1% noise, for Example 2.
p(%)parameter RMSE
λ r(t) u(0, t) u(1, t) ux(0, t) ux(1, t)0 λ=1.0E-5 1.63E-3 1.40E-5 9.88E-6 1.77E-5 1.54E-5
0.1 λdis=2.7E-2 2.60E-2 1.11E-3 7.93E-4 1.48E-3 1.02E-30.5 λdis=3.1E-1 1.02E-1 4.51E-3 3.11E-3 6.84E-3 4.21E-31 λdis=1.1 1.28E-1 8.10E-3 5.84E-3 1.09E-2 7.61E-3
10−4
10−3
10−2
10−1
100
101
102
103
10−2
10−1
100
ε = 0.023λ = 2.7E−2
ε = 0.11λ = 3.1E−1
ε = 0.24λ = 1.1
λ
‖X
r ¯λ−
y ¯ǫ‖
(a)
0 0.2 0.4 0.6 0.8 11
1.05
1.1
1.15
1.2
1.25
1.3
1.35
1.4
1.45
1.5
t
r(t
)
(b)
Figure 2.24: (a) The discrepancy principle curve, and the analytical (—–) and nu-merical results of (b)r(t) obtained using the first-order Tikhonov regularisation forp ∈ 0( ), 0.1(− · −), 0.5(· · · ), 1(− − −)% noise with the regularisation pa-rametersλdis given in Table 2.8, for Example 2.
2.5 Conclusions
In this chapter, inverse problems with nonlocal boundary conditions have been inves-
tigated in order to find the time-dependent heat source and the temperature entering
Chapter 2. 59
equation (2.1). Three general boundary and overdetermination conditions (2.3) have
been expanded into 6 separate Cases 1–6 generating six inverse problems to solve in
the context of inverse time-dependent source identification. It was found, see Table
2.1, that most ill-posed problems are given by Cases 6 and 5 followed by Case 1. It
turns out that Cases 2–4 are less ill-posed compared to the previous ones. This is con-
sistently reflected in the accuracy and stability of the numerical results obtained with
no regularisation for both exact and noisy data. Regularisation was found essential
in all cases investigated, with the first-order regularisation performing better than the
zeroth-order one.
As we have studied in this chapter, the nonlocal boundary andoverdetermination
conditions are considered in general boundary condition form. In the next chapter, the
nonlocal boundary condition will be considered together with an integral overdetermi-
nation to find the time-dependent heat source.
Chapter 3. 60
Chapter 3
Determination of a Time-dependent
Heat Source from Nonlocal Boundary
and Integral Conditions
3.1 Introduction
In many of studies of solving inverse source problems for theheat equation, see e.g.
[15, 17, 29, 61, 63, 66–68], the overdetermination conditions were selected among
classical boundary conditions and similar conditions given at interior points located
inside the body under investigation. More general, nonlocal and integral overdetermi-
nation conditions have also been considered, see the monographs [25] and [43].
In the previous chapter, we have investigated the retrievalof the time-dependent
heat sourcer(t) and the temperatureu(x, t) in the heat equation (2.1) with the three
general boundary and overdetermination conditions (2.3).In this chapter, the inves-
tigation of finding the time-dependent heat source and the temperature for the heat
equation is still of our interest, but an integral conditionis considered as overdetermi-
nation condition.
This chapter is organised as follows. In Section 3.2, the mathematical inverse for-
61
Chapter 3. 62
mulation is described and the numerical discretisation of the problem using the BEM
is presented in Section 3.3. The TSVD and the Tikhonov regularisation are described
in Section 3.4, as procedures for overcoming the instability of the solution. Finally,
Sections 3.5 and 3.6 discuss the numerical results and highlight the conclusions of this
chapter.
3.2 Mathematical formulation
Consider the problem of finding the time-dependent heat sourcer(t) ∈ C([0, T ]) and
the temperatureu(x, t) ∈ C2,1(DT ) ∩ C1,0(DT ) which satisfy the heat conduction
equation
ut = uxx + r(t)f(x, t), (x, t) ∈ DT , (3.1)
whereL = 1 in the definition ofDT in (1.1), subject to the initial condition (1.7),
namely
u(x, 0) = u0(x), x ∈ [0, 1], (3.2)
the boundary conditions
u(0, t) = u(1, t), ux(0, t) + αu(0, t) = 0, t ∈ [0, T ], (3.3)
and the energy/mass specification
∫ 1
0
u(x, t)dx = E(t), t ∈ [0, T ], (3.4)
whereα 6= 0 is a given constant andf , u0, E are given functions. The first periodic-
ity condition in (3.3) is nonlocal, whilst (3.4) specifies anintegral specification of the
energy of the thermal system. The nature of the boundary conditions (3.3) in mathe-
matical biology is demonstrated in [41], whilst the prescription of the energy, or mass,
(3.4), is encountered in heat transfer applications, [10, 22].
The unique solvability of the inverse problem (3.1)–(3.4) has been established in
Chapter 3. 63
[18], as given by the following theorem.
Theorem 3.2.1 Let the following conditions be satisfied:
(A1) u0(x) ∈ C2[0, 1]; u0(0) = u0(1), u′0(0) + αu0(0) = 0;
(A2) E(t) ∈ C1[0, T ]; E(0) =∫ 1
0u0(x)dx;
(A3) f(x, t) ∈ C(DT ); f(·, t) ∈ C2[0, 1], ∀t ∈ [0, T ]; f(0, t) = f(1, t),
fx(0, t) + αf(0, t) = 0; and∫ 1
0f(x, t)dx 6= 0, ∀t ∈ [0, T ].
Then the inverse problem(3.1)–(3.4)has a unique solution(r(t), u(x, t)) ∈ C[0, T ]×(C2,1(DT ) ∩ C1,0(DT )).
Although the inverse problem (3.1)–(3.4) is uniquely solvable, it is still ill-posed. In
the next section we will demonstrate how the BEM discretising numerically, the heat
equation (3.1) can be used together with the regularisation, to be described in Section
3.4 either the TSVD or the Tikhonov regularisation, in orderto obtain stable solutions.
3.3 The boundary element method (BEM)
In this section, we use the BEM to discretise the heat equation (3.1). As introduced in
Section 1.3, the use of BEM results in the boundary integral equation
η(x)u(x, t) =
∫ t
0
[
G(x, t, ξ, τ)∂u
∂n(ξ)(ξ, τ)− u(ξ, τ)
∂G
∂n(ξ)(x, t, ξ, τ)
]
ξ∈0,1
dτ
+
∫ 1
0
G(x, t, y, 0)u(y, 0) dy+
∫ 1
0
∫ t
0
G(x, t, y, τ)r(τ)f(y, τ) dτdy,
(x, t) ∈ [0, 1]× (0, T ]. (3.5)
Chapter 3. 64
Using the same discretisation as described in Section 2.3, we obtain
η(x)u(x, t) =
N∑
j=1
[A0j(x, t)q0j + ALj(x, t)qLj − B0j(x, t)h0j − BLj(x, t)hLj]
+
N0∑
k=1
Ck(x, t)u0,k + d(x, t), (3.6)
where source functionsr(t) andf(x, t) are assumed to be piecewise constant approx-
imations as defined in (2.21). Then, the double integral termis approximated as
d(x, t) =
∫ t
0
r(τ)
(∫ 1
0
G(x, t, y, τ)f(y, τ) dy
)
dτ =
N∑
j=1
Dj(x, t)rj ,
where
Dj(x, t) =
∫ tj
tj−1
∫ 1
0
G(x, t, y, τ)f(y, tj)dydτ =
∫ 1
0
f(y, tj)Ayj(x, t) dy,
is calculated using the midpoint integration rule. Here, the integral equation (3.6) can
be written as
η(x)u(x, t) =N∑
j=1
[A0j(x, t)q0j + ALj(x, t)qLj − B0j(x, t)h0j − BLj(x, t)hLj]
+
N0∑
k=1
Ck(x, t)u0,k +N∑
j=1
Dj(x, t)rj. (3.7)
By applying (3.7) at the boundary nodes(0, ti) and(1, ti) for i = 1, N , we obtain the
system of2N equations
Aq¯−Bh
¯+ Cu
¯0+Dr
¯= 0
¯, (3.8)
Chapter 3. 65
where all matrices and vectors are defined the same as in (2.25). From the boundary
conditions (3.3), we can express the boundary temperature h¯
as
h¯=
h0j
hLj
=
u(0, tj)
u(1, tj)
=
− 1αux(0, tj)
u(0, tj)
=
− 1αux(0, tj)
− 1αux(0, tj)
=1
α
q0j
q0j
.
Then, the system of equations (3.8) can be rewritten as
(
A− 1
α(B +B∗)
)
q¯+ Cu
¯0+Dr
¯= 0
¯, (3.9)
whereB∗ =
BLj(0, ti) −BLj(0, ti)
BLj(1, ti) +12δij −BLj(1, ti)− 1
2δij
2N×2N
. Then, we can express
the flux q¯
as
q¯= −
(
A− 1
α(B +B∗)
)−1
(Cu¯0
+Dr¯) . (3.10)
We now discretise the integral expression in (3.4), via the midpoint numerical in-
tegral approximation, as
∫ 1
0
u(x, t) dx =1
N0
N0∑
k=1
u(xk, t).
Applying this att = ti for i = 1, N , we obtain
1
N0
N0∑
k=1
u(xk, ti) =
∫ 1
0
u(x, ti) dx = E(ti) =: ei for i = 1, N. (3.11)
Using the integral equation (3.5), as before, equation (3.11) results in the system ofN
equations
1
N0
N0∑
k=1
[(
A(1)k − 1
α(B
(1)k +B
(1)k
∗)
)
q¯+ C
(1)k u
¯0+D
(1)k r
¯
]
= E¯, (3.12)
Chapter 3. 66
where
A(1)k =
[
A0j(xk, ti) ALj(xk, ti)]
N×2N, B
(1)k =
[
B0j(xk, ti) BLj(xk, ti)]
N×2N,
B(1)k
∗=[
BLj(xk, ti) −BLj(xk, ti)]
N×2N, C
(1)k =
[
Cl(xk, ti)]
N×N0
,
and D(1)k =
[
Dj(xk, ti)]
N×N, E
¯=[
ei
]
N, for k, l = 1, N0, i, j = 1, N.
Substituting the expression (3.10) into (3.12), we obtain asystem ofN ×N linear
equations as follows:
Xr¯= y
¯, (3.13)
where X =1
N0
N0∑
k=1
[
−(
A(1)k − 1
α(B
(1)k +B
(1)k
∗)
)(
A− 1
α(B +B∗)
)−1
D +D(1)k
]
and y¯= E
¯+
1
N0
N0∑
k=1
[
(
A(1)k − 1
α(B
(1)k +B
(1)k
∗)
)(
A− 1
α(B +B∗)
)−1
C − C(1)k
]
u¯0
.
Since the problem is ill-posed, then the system of equations(3.13) is ill-conditioned.
In the next section, we will deal with this ill-conditioningusing regularisation in order
to obtain a stable solution.
3.4 Regularisation
As we have mentioned before, the inverse time-dependent heat source problem (3.1)–
(3.4) has a unique solution. However this inverse problem isstill ill-posed since small
errors into the energy measurement (3.4) result in large errors in the solution r¯. In
order to model this, we investigate the stability of the numerical solution with respect
to noise in the energy data (3.4), defined as
E¯ǫ = E
¯+ random(′Normal′, 0, σ, 1, N), (3.14)
Chapter 3. 67
with the standard deviation computed by
σ = p× maxt∈[0,T ]
|E(t)|. (3.15)
When the noise is present the right-hand side of equation (3.13) is contaminated by
some noiseǫ,
‖y¯ǫ − y
¯‖ ≈ ǫ, (3.16)
therefore we have to solve the following system of linear equations instead of (3.13):
Xr¯= y
¯ǫ, (3.17)
then the inverse solution r¯= X−1y
¯ǫ of (3.13) will be a poor unstable approximation to
the exact solution. This instability can be overcome by employing the regularisation
method, and this study, we utilise either the TSVD or the Tikhonov regularisation
methods to solve the linear and ill-conditioned system of equations (3.17).
When the noise is present, we employ the TSVD procedure, as described in Sec-
tion 1.6.1 by using the[U,Σ,V]=svds(X,Nt) command in MATLAB where theNt is the
truncated level indicated by theL-curve method or the discrepancy principle. Alterna-
tively, the regularisation namely the Tikhonov regularisation is another way to obtain
a stable solution, as described in Subsection 1.6.2, yielding the regularised solution
(1.34).
3.5 Numerical examples and discussion
In this section, we present a couple of benchmark test examples to illustrate the accu-
racy and stability of the BEM combined with the Tikhonov regularisation technique
presented in Section 3.4. In order to illustrate the accuracy of the numerical results,
the RMSE defined in (2.49) is also used here.
Chapter 3. 68
3.5.1 Example 1
In the first example, we consider the inverse problem (3.1)–(3.4) withT = 1, the input
data
u(x, 0) = u0(x) = 1 + x− x2, f(x, t) = (3 + x− x2)e−t,∫ 1
0
u(x, t) dx = E(t) =7et
6,
(3.18)
and subject to the boundary condition (3.3) withα = −1. Then the analytical solution
of the inverse problem (3.1)–(3.4) is given by
u(x, t) = (1 + x− x2)et, r(t) = e2t. (3.19)
The normalised singular values of matrixX in equation (3.13) forN0 = N ∈20, 40, 80 are shown in Figure 3.1. The corresponding condition numbers of matrix
X are 248, 780, and 2393 forN0 = N ∈ 20, 40, 80, respectively. These values
indicate that the system of equations (3.13) is mildly ill-conditioned [42].
0 10 20 30 40 50 60 70 8010
−4
10−3
10−2
10−1
100
Nor
mal
ised
Sin
gula
r V
alue
s
N
Figure 3.1: The normalised singular values of matrixX for N0 = N ∈ 20 (− ·−), 40 (· · · ), 80 (−−−), for Example 1.
In what follows, we present numerical results obtained withN0 = N = 40. We
consider first the case of exact data, i.e. there is no noise inthe input data (3.4).
Figure 3.2(a) shows the analytical and numerical results ofr(t) from which it can be
Chapter 3. 69
seen that the numerical solution is very accurate. Figure 3.2(b) shows the numerical
solution ofu(0, t) in comparison with the analytical valueu(0, t) = et. The same
very good agreement between the numerical and analytical solutions is recorded. We
do not present the results for the flux atx = 1 since these were found to be equal to
the negative of the boundary temperature atx = 0. This is to be expected since from
(3.18) we haveu0(x) = u0(1 − x), f(x, t) = f(1 − x, t) and it can easily be verified
that from (3.8), using (3.3) withα = −1, it follows that q¯L
= −h¯0
. Furthermore, note
thatλ = 0, i.e. no regularisation, was found necessary forN0 = N = 40 and exact
input data (3.4).
0 0.2 0.4 0.6 0.8 11
2
3
4
5
6
7
8
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(0
,t)
(b)
Figure 3.2: The analytical (—–) and numerical (−·−) results of (a)r(t) and (b)u(0, t)for exact data andλ = 0, for Example 1.
Next, we investigate the stability of the numerical solution with respect to noise
in the energy data (3.4), we add noise to the right-hand side of the oversdetermina-
tion condition (3.4),Eǫ, as generated in (3.14) with the standard deviation given by
σ =7ep
6. The contaminated data input withp = 1% is firstly inverted by using the
Gaussian elimination procedure as shown in Figure 3.3. Although from Figure 3.3(b)
the numerical solution foru(0, t) seems to remain stable and accurate, whereas Figure
3.3(a) shows that the numerical solution forr(t), with no regularisation, is unstable,
i.e. highly oscillatory and unbounded. This phenomenon is to be expected since the
inverse problem under investigation is ill-posed with the smallest normalised singular
Chapter 3. 70
value forN = 40 being ofO(103). Moreover, similar unstable results are obtained by
using the untruncated SVD or withλ = 0 in regularised solution (1.34), instead of the
Gaussian elimination method.
0 0.2 0.4 0.6 0.8 1−20
−15
−10
−5
0
5
10
15
20
25
30
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(0
,t)
(b)
Figure 3.3: The analytical (—–) and numerical (−·−) results of (a)r(t) and (b)u(0, t)for p = 1% noisy data andλ = 0, for Example 1.
In order to avoid the over amplification of noise and maintainthe stability of the
results when the noise is presented, we employ the regularisation of either the TSVD
or the Tikhonov regularisation of orders zero, one and two. Furthermore, theL-curve
method and the discrepancy principle are investigated to indicate the truncation level
Nt for using TSVD, and choose the regularisation parameterλ for using the Tikhonov
regularisation. Firstly, we consider the BEM together withthe TSVD in the case of
p ∈ 1, 3, 5%. Initially, we have tried theL-curve criterion which should select the
truncated numberNt at the corner ofL-curve graph. Figure 3.4 displays theL-curves
for various percentages of noisep ∈ 1, 3, 5%. From this figure, especially on the nor-
mal scale one can see thatL-shaped curves appear and their corners indicate the level
of truncationNt in the following intervals9, . . . , 32, 12, . . . , 24 and10, . . . , 17for p ∈ 1, 3, 5%, respectively. Alternatively, one can choose the truncation level by
the more rigorous discrepancy principle which uses the knowledge of the level of noise
as in (3.16), i.e.ǫ = ‖y¯ǫ − y
¯‖ = ‖E
¯ǫ − E‖. Figure 3.5(a) represents the discrepancy
principle to selectNt for p ∈ 1, 3, 5%. These truncated numbersNt are obtained
Chapter 3. 71
within the intervals which have been recommended by theL-curve criterion of Figure
3.4. Besides, the numerical results forr(t) andu(0, t) are illustrated in Figures 3.5(b)
and 3.5(c), respectively, and the RMSE for these results aregiven in Table 3.1. From
Table 3.1 it can be remarked that the accuracy of the numerical results foru(0, t) im-
proves as the percentage of noise decreases; however, the convergence of the numerical
solution forr(t) towards the analytical solution is non-monotonic asp decreases (to
zero), see also Figure 3.5(b). Further, from Figure 3.5(c),it can be seen that the nu-
merical results foru(0, t) are stable, although by comparing with Figure 3.3(b) it can
be seen that forp = 1% noise the results with untruncation, i.e. no regularisation, are
slightly more accurate than the results withNt = 15. This is somewhat to be expected
since the pair of solutions(u(x, t), r(t)) of the inverse problem is stable in the temper-
atureu(x, t) component, but unstable in the heat sourcer(t) component. This latter
instability in r(t) which has already been illustrated in Figure 3.3(b), is alleviated in
Figure 3.5(b) which shows the numerical results with regularisation.
Alternatively for comparison, the zeroth-order Tikhonov regularisation (ZOTR) is
also utilised here. Initially, we have tried theL-curve plot of the residual‖Xr¯λ
−y¯ǫ‖ versus the norm of the solution‖r
¯λ‖, but we have found that anL-corner could
not be clearly identified. This is to be somewhat expected because theL-curve is a
heuristic method which is not always convergent [58]. Another optional method, the
more rigorous discrepancy principle which uses the knowledge of the level of noiseǫ
is recommended.
Figure 3.6(a) shows the discrepancy principle curves, for various percentages of
noisep ∈ 1, 3, 5% where the intersection of the residual curve with the horizontal
level noisy lineǫ yields the value ofλdis. The RMSE for the output solutions ofr(t)
andu(0, t) obtained with these values ofλdis are given in Table 3.1 and the numerical
results are illustrated in Figures 3.6(b) and 3.6(c). From this table it can be seen that the
ZOTR slightly outperforms the TSVD. As it can be seen from Figures 3.6(b) and 3.6(c)
the numerical results obtained by the ZOTR are stable and in good agreement with the
analytical solution. Although Figure 3.6(b) represents a significant improvement over
Chapter 3. 72
0.05 0.1 1 1010
12
14
16
18
20
22
24
26
28
Nt=1
Nt=15
Nt=20,...,32
Nt=19
Nt=39
‖Xr¯Nt
− y¯
ǫ‖
‖r ¯N
t‖
(a)
0 1 2 3 4 5 610
20
30
40
50
60
70
80
Nt=1
Nt=9,...,32
Nt=40
‖Xr¯Nt
− y¯
ǫ‖
‖r ¯N
t‖
(b)
100
10
20
30
40
50
Nt=1
Nt=10,11
Nt=9Nt=12,...,14
Nt=39
‖Xr¯Nt
− y¯
ǫ‖
‖r ¯N
t‖
Nt=15,...,24
(c)
0 1 2 3 4 5 60
20
40
60
80
100
120
140
160
Nt=1
Nt=40
‖Xr¯Nt
− y¯
ǫ‖
‖r ¯N
t‖
Nt=12,...,24
Nt=9
(d)
10−3
10−2
10−1
100
101
10
20
40
60
80
100
120
Nt=1
Nt=7,...,9
Nt=7
Nt=10,...,17
Nt=18,...,21
Nt=39
‖Xr¯Nt
− y¯
ǫ‖
‖r ¯N
t‖
(e)
0 1 2 3 4 510
20
30
40
50
60
70
80
90
100
110
Nt=1
Nt=7,...,9
Nt=10,...,17
Nt=18,...,21
Nt=40
‖Xr¯Nt
− y¯
ǫ‖
‖r ¯N
t‖
(f)
Figure 3.4: TheL-curve on (a), (c), (e) log-log scale and (b), (d), (f) normalscale,obtained using TSVD forp ∈ 1(top), 3(middle), 5(bottom)% noise, for Example1.
Chapter 3. 73
0 5 10 15 20 25 30 35 4010
−3
10−2
10−1
100
101
ε = 0.17Nt = 15
ε = 0.56Nt = 9
ε = 0.91Nt = 10
Nt
‖X
r ¯Nt−
y ¯ǫ‖
(a)
0 0.2 0.4 0.6 0.8 1−1
0
1
2
3
4
5
6
7
8
t
r(t
)
(b)
0 0.2 0.4 0.6 0.8 10.8
1
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(0
,t)
(c)
Figure 3.5: (a) The discrepancy principle curve, and the analytical (—–) and nu-merical results of (b)r(t), and (c)u(0, t) obtained using the TSVD forp ∈ 1(− ·−), 3(· · · ), 5(− − −)% noise at the truncation levelNt given in Table 3.1, for Ex-ample 1.
Figure 3.3(a) the numerical results are still oscillatory.In order to further improve on
the accuracy and stability of the numerical results of Figure 3.6(b), next we employ the
regularisation with the higher-order Tikhonov regularisation of orders one and two.
Figures 3.7 and 3.8 show the analogous numerical results to Figure 3.6, but for
the first- and second-order Tikhonov regularisation, denoted FOTR and SOTR, respec-
tively. The corresponding values of the regularisation parametersλdis chosen accord-
ing to the discrepancy principle are given in Table 3.1. By comparing Figures 3.6–3.8
and Table 3.1, it can be seen that the use of higher-order regularisation improves the
Chapter 3. 74
10−6
10−5
10−4
10−3
10−2
10−1
100
101
10−2
10−1
100
101
102
ε = 0.17λ = 8E−4
ε = 0.56λ = 4E−3
ε = 0.91λ = 5E−3
λ
‖X
r ¯λ−
y ¯ǫ‖
(a)
0 0.2 0.4 0.6 0.8 10
1
2
3
4
5
6
7
8
t
r(t
)
(b)
0 0.2 0.4 0.6 0.8 10.8
1
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(0
,t)
(c)
Figure 3.6: (a) The discrepancy principle curve, and the analytical (—–) and nu-merical results of (b)r(t), and (c)u(0, t) obtained using the ZOTR forp ∈ 1(− ·−), 3(· · · ), 5(−−−)% noise with the regularisation parametersλdis given in Table3.1, for Example 1.
accuracy and stability of the numerical results. This is to be expected since the analyt-
ical solution (3.19) to be retrieved is infinitely differentiable. However, if less smooth
sources are attempted to be retrieved, the use of higher-order regularisation does not
necessarily improve the accuracy of the numerical results,as can be seen from the next
example.
Chapter 3. 75
Table 3.1: The RMSE for the TSVD, ZOTR, FOTR, SOTR forp ∈ 0, 1, 3, 5% noise,for Example 1.
Regularisation p(%) parameterRMSE
r(t) u(0, t)
-0 0 8.1E-3 6.6E-41% 0 11.66 0.023
TSVD1% Nt=15 1.080 2.47E-23% Nt=9 1.443 6.03E-25% Nt=10 1.521 7.25E-2
ZOTR1% λdis=8E-4 0.878 2.40E-23% λdis=4E-3 1.324 6.22E-25% λdis=5E-3 1.280 9.90E-2
FOTR1% λdis=0.05 0.246 1.17E-23% λdis=0.45 0.435 3.14E-25% λdis=0.64 0.595 5.81E-2
SOTR1% λdis=5.2 0.113 6.86E-33% λdis=81 0.198 2.67E-25% λdis=324 0.473 5.48E-2
3.5.2 Example 2
The previous Example 1 involved retrieving a smooth source function given byr(t) =
e2t. In this example, we are considering the BEM combined only with the Tikhonov
regularisation as we found in the previous example that the ZOTR slightly outperforms
the TSVD. Consider a more severe discontinuous test function given by [63],
r(t) =
−1, t ∈ [0, 0.25),
1, t ∈ [0.25, 0.5),
−1, t ∈ [0.5, 0.75),
1, t ∈ [0.75, 1].
(3.20)
We also takeT = 1,α = −1, u0 = 0, andf = 1. Since the inverse problem (3.1)–(3.4)
does not have an analytical solution available for the temperatureu(x, t), the input en-
ergy data (3.4) is numerically simulated by solving first thedirect problem (3.1)–(3.2)
Chapter 3. 76
10−6
10−5
10−4
10−3
10−2
10−1
100
101
102
103
10−2
10−1
100
101
ε = 0.17λ = 0.05
ε = 0.56λ = 0.45
ε = 0.91λ = 0.64
λ
‖X
r ¯λ−
y ¯ǫ‖
(a)
0 0.2 0.4 0.6 0.8 11
2
3
4
5
6
7
8
t
r(t
)
(b)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(0
,t)
(c)
Figure 3.7: (a) The discrepancy principle curve, and the analytical (—–) and nu-merical results of (b)r(t), and (c)u(0, t) obtained using the FOTR forp ∈ 1(− ·−), 3(· · · ), 5(−−−)% noise with the regularisation parametersλdis given in Table3.1, for Example 1.
with r given by (3.20). The numerical results forE(t) andu(0, t) are shown in Fig-
ures 3.9(a) and 3.9(b), respectively, for various numbers of boundary elements/cells
N = N0 ∈ 20, 40, 80. From Figure 3.9 it can be seen that the numerical solution is
convergent as the number of boundary elements increases. Also, there is little differ-
ence between the numerical results obtained with the various mesh sizes showing that
the independence of the mesh has been achieved. We can therefore take the numerical
results forE(t), simulated from the direct problem withN = N0 = 40 and shown in
Figure 3.9(a), as our input (3.4) in the inverse problem (3.1)–(3.4). Furthermore, in
Chapter 3. 77
10−3
10−2
10−1
100
101
102
103
104
105
106
0.1
0.2
0.4
0.6
0.8
1
ε = 0.17 λ = 5.2
ε = 0.56 λ = 81.9
ε = 0.91 λ = 324
λ
‖X
r ¯λ−
y ¯ǫ‖
(a)
0 0.2 0.4 0.6 0.8 10
1
2
3
4
5
6
7
8
t
r(t
)
(b)
0 0.2 0.4 0.6 0.8 11
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
t
u(0
,t)
(c)
Figure 3.8: (a) The discrepancy principle curve, and the analytical (—–) and nu-merical results of (b)r(t), and (c)u(0, t) obtained using the SOTR forp ∈ 1(− ·−), 3(· · · ), 5(−−−)% noise with the regularisation parametersλdis given in Table3.1, for Example 1.
order to avoid committing an inverse crime [32], in the inverse problem the number of
cells is takenN0 = 30 (different than 40) while the number of boundary elements is
kept the sameN = 40.
First, Figures 3.10(a) and 3.10(b) show the numerical results for r(t) andu(0, t),
respectively, for no noise in the input data (3.4) obtained with no regularisation, i.e.
λ = 0. The exact solution (3.20) forr(t) is also included, and the numerical solution
for u(0, t) from Figure 3.9(b) withN = N0 = 40 is also referred to as ‘analytical’.
From Figure 3.10 it can be seen that the agreement between thenumerical and the
Chapter 3. 78
0 0.2 0.4 0.6 0.8 1−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
t
E(t
)
(a)
0 0.2 0.4 0.6 0.8 1−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
t
u(0
,t)
(b)
Figure 3.9: The numerical results of (a)E(t) and (b)u(0, t) obtained by solving thedirect problem withN0 = N ∈ 20 (− · −), 40 (· · · ), 80 (−−−), for Example 2.
0 0.2 0.4 0.6 0.8 1−1.5
−1
−0.5
0
0.5
1
1.5
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 1−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
t
u(0
,t)
(b)
Figure 3.10: The analytical (—–) and numerical(− · −) results of (a)r(t) and (b)u(0, t) for exact data andλ = 0, for Example 2.
analytical solutions is excellent. Next we add noise to the input data (3.4) numerically
simulated in Figure 3.9(a). This is generated as in (3.14) with the standard deviation
given byσ = 0.19p. If λ = 0, the numerically retrieved results forr(t) were found
highly unbounded and oscillatory and therefore, they are not presented. The numerical
results for the discrepancy principle,r(t) andu(0, t) for various amounts of noise and
regularisation of various orders zero, one, and two are shown in Figures 3.11–3.13, re-
spectively, and Table 3.2. The use of higher-order regularisation imposes higher-order
Chapter 3. 79
Table 3.2: The RMSE for the ZOTR, FOTR, SOTR forp ∈ 1, 3, 5% noise andN = 40,N0 = 30, for Example 2.
Regularisation p(%)parameter RMSEλdis r(t) u(0, t)
ZOTR1% 8.7E-5 0.195 1.90E-33% 3.3E-4 0.284 4.66E-35% 4.3E-4 0.287 5.80E-3
FOTR1% 9.9E-5 0.211 1.98E-33% 8.1E-4 0.289 4.38E-35% 7.8E-4 0.280 4.69E-3
SOTR1% 7.7E-5 0.223 2.06E-33% 9.2E-4 0.292 4.48E-35% 7.3E-4 0.279 4.48E-3
smoothness of the desired output hence, since the solution (3.20) is discontinuous,
more accurate results are obtained with the ZOTR, whilst theFOTR and the SOTR,
although they achieve stability, they slightly oversmooththe retrieved solution.
3.6 Conclusions
In this chapter, the inverse problem of finding the time-dependent heat source together
with the temperature of heat equation under non-local boundary and integral overdeter-
mination conditions has been investigated. The general ill-posed problem, a numerical
method based on the BEM combined with either the TSVD or the Tikhonov regularisa-
tion has been proposed. The TSVD has been truncated at the optimal truncation level
given by theL-curve criterion and the discrepancy principle. Whereas the discrep-
ancy principle for choosing the regularisation parameter with three orders of Tikhonov
regularisation have also been employed. The numerical results were found to be ac-
curate and stable. These features of the numerical solutionincrease with decreasing
the amount of noise included in the input data and with increasing the order of reg-
ularisation for smooth sources. However, as expected, non-smooth sources are more
accurately reconstructed by lower-order regularisation.
Chapter 3. 80
10−6
10−5
10−4
10−3
10−2
10−1
100
101
10−3
10−2
10−1
100
ε = 0.011λ = 8.7E−5
ε = 0.037λ = 3.3E−4
ε = 0.059λ = 4.3E−4
λ
‖X
r ¯λ−
y ¯ǫ‖
(a)
0 0.2 0.4 0.6 0.8 1−1.5
−1
−0.5
0
0.5
1
1.5
t
r(t
)
(b)
0 0.2 0.4 0.6 0.8 1−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
t
u(0
,t)
(c)
Figure 3.11: (a) The discrepancy principle curve, and the analytical (—–) and nu-merical results of (b)r(t), and (c)u(0, t) obtained using the ZOTR forp ∈ 1(− ·−), 3(· · · ), 5(−−−)% noise with the regularisation parametersλdis given in Table3.2, for Example 2.
We have studied the nonlocal boundary and integral overdetermination condition
for the inverse source problem for the heat equation. In the next chapter, we will subject
these conditions to a coefficient identification problem forthe bioheat equation.
Chapter 3. 81
10−6
10−5
10−4
10−3
10−2
10−1
100
101
10−3
10−2
10−1
100
ε = 0.011λ = 9.8E−5
ε = 0.037λ = 8.1E−4
ε = 0.059λ = 7.8E−4
λ
‖X
r ¯λ−
y ¯ǫ‖
(a)
0 0.2 0.4 0.6 0.8 1−1.5
−1
−0.5
0
0.5
1
1.5
t
r(t
)
(b)
0 0.2 0.4 0.6 0.8 1−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
t
u(0
,t)
(c)
Figure 3.12: (a) The discrepancy principle curve, and the analytical (—–) and nu-merical results of (b)r(t), and (c)u(0, t) obtained using the FOTR forp ∈ 1(− ·−), 3(· · · ), 5(−−−)% noise with the regularisation parametersλdis given in Table3.2, for Example 2.
Chapter 3. 82
10−6
10−5
10−4
10−3
10−2
10−1
100
101
10−3
10−2
10−1
100
ε = 0.011λ = 7.7E−5
ε = 0.037λ = 9.2E−4
ε = 0.059λ = 7.3E−4
λ
‖X
r ¯λ−
y ¯ǫ‖
(a)
0 0.2 0.4 0.6 0.8 1−1.5
−1
−0.5
0
0.5
1
1.5
t
r(t
)
(b)
0 0.2 0.4 0.6 0.8 1−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
t
u(0
,t)
(c)
Figure 3.13: (a) The discrepancy principle curve, and the analytical (—–) and nu-merical results of (b)r(t), and (c)u(0, t) obtained using the SOTR forp ∈ 1(− ·−), 3(· · · ), 5(−−−)% noise with the regularisation parametersλdis given in Table3.2, for Example 2.
Chapter 4
Determination of a Time-dependent
Coefficient in the Bioheat Equation
4.1 Introduction
The bioheat equation establishes a mathematical connection between the tissue tem-
perature and the arterial blood perfusion which are the dominant components in human
physiology, see [55]. It involves a blood perfusion coefficient whose determination is
of much interest, [53].
In the previous chapter, we have considered the identification of the time-dependent
heat sourcer(t) and the temperatureu(x, t) in the heat equation (3.1) with nonlocal
boundary and integral conditions. In this chapter, we aim tofind the time-dependent
blood perfusion coefficient functionP (t) and the temperature of the tissueu(x, t) en-
tering the bioheat equation (4.1) below. The initial (3.2),nonlocal boundary (3.3) and
integral (3.4) conditions are the same as in the previous chapter. We mention that time-
dependent coefficient identification problems with nonlocal boundary and/or integral
overdetermination conditions have recently attracted revitalising interest, e.g. the re-
construction of a time-dependent diffusivity [27], a bloodperfusion coefficient [23], or
a heat source [18, 24].
83
Chapter 4. 84
The inverse problem investigated in this chapter has already been proved to be
uniquely solvable in [33], but no numerical reconstructionhas been attempted. There-
fore, the purpose of this chapter is to devise a numerical stable method for obtaining
the solution of the inverse problem.
4.2 Mathematical formulation
Let us consider the inverse problem consisting of finding thetime-dependent blood per-
fusion coefficient functionP (t) ∈ C[0, T ] and the temperature of the tissueu(x, t) ∈C2,1(DT )∩C1,0(DT ) satisfying the one-dimensional time-dependent bioheat equation,
[33, 54],
ut(x, t) = uxx(x, t)− P (t)u(x, t) + f(x, t), (x, t) ∈ DT , (4.1)
wheref is a known heat source term andL = 1 in the definition ofDT in (1.1), subject
to the following initial, boundary and overdetermination conditions:
u(x, 0) = u0(x), x ∈ [0, 1], (4.2)
u(0, t) = u(1, t), ux(0, t) + αu(0, t) = 0, t ∈ [0, T ], (4.3)∫ 1
0
u(x, t)dx = E(t), t ∈ [0, T ], (4.4)
where the functionu0 is given and it denotes the initial temperature,α is a given
constant heat transfer coefficient, andE represents the mass or energy of the system.
Note that the nonlocal periodic boundary condition (4.3) isencountered in biological
applications, [41], whilst the mass/energy specification (4.4) models processes related
to particle diffusion in turbulent plasma, [22], or heat conduction, [10]. The physical
constraint that the blood perfusionP (t) is positive can also be imposed, [37].
Note that the caseα = 0 has been dealt with in [24]. Herein, we consider the case
α 6= 0 whose unique solvability and local continuous dependence of the solution upon
Chapter 4. 85
the data of the inverse problem (4.1)–(4.4) have been established in [33]. Moreover,
the continuous dependence of the solution upon the data alsowas established, [21].
Consider now the following transformation, [11],
v(x, t) = r(t)u(x, t), r(t) = exp
(∫ t
0
P (τ) dτ
)
. (4.5)
Then the inverse problem (4.1)–(4.3) becomes
vt = vxx + r(t)f(x, t), (x, t) ∈ DT , (4.6)
v(x, 0) = u0(x), x ∈ [0, t], (4.7)
v(0, t) = v(1, t), vx(0, t) + αv(0, t) = 0, t ∈ [0, T ], (4.8)
with the transformed integral condition
∫ 1
0
v(x, t)dx = E(t)r(t), t ∈ [0, T ]. (4.9)
We also have thatr ∈ C1[0, T ], r(0) = 1, r(t) > 0, for t ∈ [0, T ]. Solving the inverse
problem (4.6)–(4.9) for the solution pair(r(t), v(x, t)) yields afterwards the solution
pair (P (t), u(x, t)) for the inverse problem (4.1)–(4.4) as given by
P (t) =r′(t)
r(t)and u(x, t) =
v(x, t)
r(t), (x, t) ∈ DT . (4.10)
From equation (4.10) one can observe that the ill-posednessof the inverse problem
consists of the numerical differentiation of the noisy function r(t) which would need
regularisation.
4.3 The boundary element method (BEM)
In this section, we apply the BEM to the one-dimensional inverse problem (4.6)–(4.9),
in order to approximate the solution(r(t), v(x, t)) which in turn, via (4.10), leads to
Chapter 4. 86
the original solution(P (t), u(x, t)) of the inverse problem (4.1)–(4.4). Utilising the
BEM is classical with the use of the fundamental solution forthe heat equation and
Green’s identities. As introduced in Section 1.3, we then obtain the boundary integral
equation
η(x)v(x, t) =
∫ t
0
[
G(x, t, ξ, τ)∂v
∂n(ξ)(ξ, τ)− v(ξ, τ)
∂G
∂n(ξ)(x, t, ξ, τ)
]
ξ∈0,1
dτ
+
∫ 1
0
G(x, t, y, 0)v(y, 0) dy+
∫ 1
0
∫ T
0
G(x, t, y, τ)r(τ)f(y, τ) dτdy,
(x, t) ∈ [0, 1]× (0, T ]. (4.11)
Using the same discretisation as in Chapters 1–3, then the integral equation above can
be approximate as
η(x)v(x, t) =
N∑
j=1
[A0j(x, t)q0j + ALj(x, t)qLj − B0j(x, t)h0j − BLj(x, t)hLj ]
+
N0∑
k=1
Ck(x, t)u0,k +N∑
j=1
Dj(x, t)rj . (4.12)
By applying the boundary condition (4.3), we obtain the system of2N linear equations,
the same as in (3.9),
(
A− 1
α(B +B∗)
)
q¯+ Cu
¯0+Dr
¯= 0
¯. (4.13)
The transformed integral condition (4.9) can also be expressed, via the midpoint nu-
merical integral approximation, as
1
N0
N0∑
k=1
v(xk, ti),=
∫ 1
0
v(x, ti) dx = E(ti)r(ti) = eiri, i = 1, N. (4.14)
Chapter 4. 87
By using the integral equation (4.12), via (4.13), and (4.14) as the same in Chapter 3
yields
1
N0
N0∑
k=1
[(
A(1)k − 1
α(B
(1)k +B
(1)∗
k )
)
q¯+ C
(1)k u
¯0+D
(1)k r
¯
]
= Er¯, (4.15)
whereE = diag(e1, eN ). Eliminating q¯
from (4.13) and (4.15) yields a linear system
of N equations
Xr¯= y
¯, (4.16)
with N unknowns, where
X =1
N0
N0∑
k=1
[
−(
A(1)k − 1
α(B
(1)k +B
(1)k
∗)
)(
A− 1
α(B +B∗)
)−1
D +D(1)k
]
− E,
y¯=
1
N0
N0∑
k=1
[
(
A(1)k − 1
α(B
(1)k +B
(1)k
∗)
)(
A− 1
α(B +B∗)
)−1
C − C(1)k
]
u¯0.
As we have mention before, this inverse problem is ill-posed, then we need to
employ the regularisation in order to retrieve the stability of the solution which will be
present in the next section.
4.4 Regularisation
In practical measurements, the data (4.4) is usually contaminated with noise. In order
to model this, we perturb (4.4) with random noise as defined in(3.14), i.e. E¯ǫ =
E¯+ ǫ. Then, from the contamination, it means that the left-hand side matrixX is
contaminated with noise, denoted byXǫ, where
ǫ ≈ ‖Xǫ −X‖. (4.17)
The norm of the matrix above is defined as the square root of thesum of squares of all
its elements. Hence, instead of (4.16) we have to solve the following linear system of
Chapter 4. 88
equations:
Xǫr¯= y
¯. (4.18)
When the noise is presented, we employ the Tikhonov regularisation, as described
in Subsection 1.6.2, yielding the regularised solution
r¯λ
=(
(Xǫ)TXǫ + λRTR)−1
(Xǫ)Ty¯. (4.19)
whereλ ≥ 0 is a regularisation parameter to be prescribed andR is a second-order
derivative regularisation matrix given by, [57],
R =1
(T/N)2
1 −2 1 0 0 .
0 1 −2 1 0 .
0 0 1 −2 1 .
. . . . . .
. (4.20)
Note that we have kept the multiple of1
(T/N)2to the regularisation matrix which
is different from the regularisation matrix introduced in Subsection 1.6.2, this is in
order to keep the scaling technique of the computation. In (4.19), the regularisation
parameter can be selected according to the GCV criterion which choosesλ > 0 as the
minimum of the GCV function, see e.g. [63] and (1.35),
GCV (λ) =‖Xǫr
¯λ− y
¯‖2
[trace(I −Xǫ((Xǫ)TXǫ + λRTR)−1(Xǫ)T)]2. (4.21)
The solution of the original inverse problem (4.1)–(4.4) can be obtained by substi-
tuting all approximate solutionsv, r, andr′ into (4.10). In order to obtain the solution
P (t), we also need to find the derivative functionr′(t) which can be approximated
using finite differences as
r′(t1) =rλ(t1)− 1
T/(2N), r′(ti) =
rλ(ti)− rλ(ti−1)
T/N, i = 2, N. (4.22)
Chapter 4. 89
In the next section, we will test numerical examples in orderto illustrate the accu-
racy and stability of the BEM combined with the regularisation technique.
4.5 Numerical examples and discussion
This section presents two benchmark test examples in order to test the accuracy and
stability of the BEM numerical procedure introduced earlier. The RMSE defined in
(2.49) is also used here to evaluate the accuracy of the numerical results.
4.5.1 Example 1
We consider a benchmark test example with the inputT = 1, α = −1 and
u(x, 0) = u0(x) = 1 + x− x2, f(x, t) = (3 + x− x2)e−t,∫ 1
0
u(x, t) dx = E(t) = 7e−t/6,(4.23)
Then the analytical solution of the problem (4.1)–(4.4) is given by
u(x, t) = (1 + x− x2)e−t, P (t) = 2, (4.24)
whilst the analytical solution for the transformed problem(4.6)–(4.9) is
v(x, t) = (1 + x− x2)et, r(t) = e2t, (4.25)
In this example, we present the numerical results obtained with a BEM mesh ofN =
N0 = 40.
We start first with the case of exact data, i.e.p = 0. The numerical results for the
unknownsr(t), u(0, t), r′(t), andP (t) obtained using the straightforward inversion
r¯= X−1y
¯, i.e. without regularisationλ = 0 in (4.19), are compared with their cor-
responding analytic valuese2t, e−t, 2e2t, and 2, in Figures 4.1(a)–4.1(d), respectively.
Chapter 4. 90
0 0.2 0.4 0.6 0.8 11
2
3
4
5
6
7
8
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 1
0.4
0.5
0.6
0.7
0.8
0.9
1
t
u(0
,t)
(b)
0 0.2 0.4 0.6 0.8 12
4
6
8
10
12
14
16
t
r′(t
)
(c)
0 0.2 0.4 0.6 0.8 11.94
1.95
1.96
1.97
1.98
1.99
2
2.01
2.02
2.03
t
P(t
)
(d)
Figure 4.1: The analytical (—–) and numerical(− · −) results of (a)r(t), (b) u(0, t),(c) r′(t), and (d)P (t) obtained using no regularisation, i.e.λ = 0, for exact data, forExample 1.
From Figure 4.1 it can be seen that all the quantities of interest are accurate.
Next we investigate the stability of the numerical solutionwith respect to some
p = 1% noise included in the input energy dataE, as mentioned in Section 4.4. The
corresponding numerical results to Figure 4.1 (for exact data) are presented in Fig-
ure 4.2 (for noisy data). In Figures 4.2(a) and 4.2(b) the numerical results obtained
for r(t) andu(0, t), respectively, are relatively accurate. However, the numerical re-
sults obtained forr′(t) andP (t) = r′(t)/r(t) shown in Figures 4.2(c) and 4.2(d),
respectively, are highly unstable. This is because the differentiation of the noisy func-
tion r(t), shown in Figure 4.2(a) with dashed line, using the finite differences (4.22)
Chapter 4. 91
0 0.2 0.4 0.6 0.8 11
2
3
4
5
6
7
8
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 1
0.4
0.5
0.6
0.7
0.8
0.9
1
t
u(0
,t)
(b)
0 0.2 0.4 0.6 0.8 1−5
0
5
10
15
20
25
t
r′(t
)
(c)
0 0.2 0.4 0.6 0.8 1−0.5
0
0.5
1
1.5
2
2.5
3
3.5
4
t
P(t
)
(d)
Figure 4.2: The analytical (—–) and numerical(− · −) results of (a)r(t), (b) u(0, t),(c) r′(t), and (d)P (t) obtained using no regularisation, i.e.λ = 0, for p = 1% noise,for Example 1.
is an unstable procedure. In order to deal with this instability one can employ the
smoothing spline regularisation of [59], but this requiresthe knowledge of the discrep-
ancy between the analytical and numerical values ofr(t) in Figure 4.2(a), which is
not available if the analytical solution is not available. We shall elaborate in apply-
ing this technique later on for Example 2. Alternatively, weemploy the second-order
Tikhonov regularised solution (4.19) with the choice of theregularisation parameter
given by the minimum point of the GCV function (4.21). This isplotted in Figure 4.3
for p = 1% noise and the minimum yields the valueλGCV = 1.25 × 10−5. With the
value ofλGCV = 1.25 × 10−5, the solution (4.19) forr(t) is plotted in Figure 4.4(a).
Chapter 4. 92
By comparing with the previous unregularised solution shown in Figure 4.2(a), one
can now see that the obtained solution forr(t) is indeed smooth. Then the process
of numerical differentiation (4.22) is permitted and a stable approximation can be ob-
tained, as shown in Figures 4.4(c) and 4.4(d). There are someinaccuracies manifested
at the end pointst = 0 and 1, but this is commonly observed elsewhere when using
other stabilising techniques such as the mollification method or, the Tikhonov regular-
isation for the Fredholm integral equation of the first kind presented in detail in [56].
In Figure 4.4 we have also included for comparative purposesthe numerical results
obtained with the optimal value, in circle line( ) of the regularisation parameter
λopt = 1.05×10−6 selected by the trail and error. The RMSE forλGCV = 1.25×10−5
andλopt = 1.05 × 10−6 for the numerical results presented in Figure 4.4(d) in com-
parison with the analytical solution, are 0.324 and 0.219, respectively. Overall from
Figure 4.4 it can be seen that there are not much differences between the numerical
results obtained with the two values of the regularisation parameter. This confirms that
the GCV criterion performs well in choosing a suitable regularisation parameter for
obtaining a stable and accurate numerical solution.
10−7
10−6
10−5
10−4
6
6.5
7
7.5
8
8.5
9
9.5
10x 10
−5
λ = 1.25×10−5
λ
GC
V(λ
)
Figure 4.3: The GCV function (4.21), obtained using the second-order Tikhonov reg-ularisation forp = 1% noise, for Example 1.
Chapter 4. 93
0 0.2 0.4 0.6 0.8 11
2
3
4
5
6
7
8
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 1
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
t
u(0
,t)
(b)
0 0.2 0.4 0.6 0.8 10
5
10
15
t
r′(t
)
(c)
0 0.2 0.4 0.6 0.8 10
0.5
1
1.5
2
2.5
t
P(t
)
(d)
Figure 4.4: The analytical (—–) and numerical results of (a)r(t), (b)u(0, t), (c) r′(t),and (d)P (t) obtained using the second-order Tikhonov regularisation with λGCV =1.25× 10−5 (− ·−) andλopt = 1.05× 10−6 ( ), for p = 1% noise, for Example 1.
4.5.2 Example 2
In the previous example, the BEM together with the second-order regularisation and
finite differences has been used to solve the inverse problemfor the bioheat equation
(4.1)–(4.4). In this example we present the BEM together with a smoothing spline
regularisation, to be utilised as another regularisation technique, for computing the
first-order derivative of a noisy function, see [59]. We consider the inverse problem
Chapter 4. 94
(4.1)–(4.4) with the inputL = T = 1, α = −1 and
u0(x) = x4(1− x)3, E(t) = 1280e−(t+t2/2),
f(x, t) = 6x2(x− 1)(7x2 − 8x+ 2)e−(t+t2/2).
(4.26)
Then the solution of the transformed inverse problem (4.6)–(4.9) is given by
v(x, t) = x4(1− x)3, r(t) = et+t2/2, (4.27)
whereas the solution of the original inverse bioheat conduction problem (4.1)–(4.4) is
u(x, t) = x4(1− x)3e−(t+t2/2), P (t) = 1 + t.
From Figure 4.2(a) of the previous example, we have noticed that the numerical
results for the perfusion coefficientP (t) are highly oscillatory and unbounded because
the numerical differentiation of a noisy function is an unstable procedure. For Example
1, we have used the second-order Tikhonov regularisation for solving the system of
equations (4.19) and this resulted in a smooth approximation for the functionr(t),
as shown in Figure 4.4(a). Alternatively, for Example 2 we investigate smoothing
a posteriorithe discrete noisy data r¯
obtained (without regularisation) like in Figure
4.2(a). This is applied as the smoothing spline technique of[59]. We are seeking
therefore a smooth functionr ∈ C1(R) with r′′ ∈ L2(R) which minimises the second-
order Tikhonov regularisation functional
IΛ(r) :=T
N
N∑
j=1
(r(tj)− rj)2 + Λ‖r′′‖2L2(R), (4.28)
whereΛ ≥ 0 is a regularisation parameter to be prescribed and
r¯= (rj)j=1,N = (Xǫ)−1y
¯, (4.29)
Chapter 4. 95
0 0.2 0.4 0.6 0.8 10.5
1
1.5
2
2.5
3
3.5
4
4.5
5
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 1−3
−2
−1
0
1
2
3x 10
−5
t
u(0
,t)
(b)
0 0.2 0.4 0.6 0.8 1−10
−5
0
5
10
15
20
25
30
t
r′(t
)
(c)
0 0.2 0.4 0.6 0.8 1−2
−1
0
1
2
3
4
5
6
t
P(t
)
(d)
Figure 4.5: The analytical (—–) and numerical results of (a)r(t), (b) u(0, t)(t), (c)r′(t), and (d)P (t) obtained using no regularisation in (4.31), i.e.Λ = 0, for exact data( ) and forp = 1% noisy data(− · −), for Example 2.
obtained from Figure 4.2(a). We further approximate the functionr using cubic splines
as
r(t) = d1 + d2t+
N∑
j=1
cj|t− tj |3, (4.30)
where the coefficients(cj)j=1,N and(dj)j=1,2 satisfy the conditions (4.31) and (4.32)
below. Inserting (4.30) into (4.28) and minimising the resulting expression with re-
spect to the coefficients(cj)j=1,N and(dj)j=1,2 yields the following system of(N +2)
Chapter 4. 96
equations with(N + 2) unknowns, [59],
rΛ(tj) + 12ΛNcj = rj , j = 1, N, (4.31)
N∑
j=1
cj =
N∑
j=1
cjtj = 0. (4.32)
By solving the system of equations (4.31) and (4.32) we obtain the coefficients(cj)j=1,N
and(dj)j=1,2, and therefore the expression for the smooth functionr, given by (4.30).
By differentiating this expression with respect tot we obtain the first-order derivative
r′(t) given by
r′(t) = d2 + 3
N∑
j=1
cj(t− tj)2sign(t− tj), (4.33)
where sign(·) is the signum function. This derivativer′(t) is presented asr′(t) in
(4.27). From (4.31), one can observe that ifΛ = 0 thenr0(ti) = r(ti) for i = 1, N .
In the case of exact data actually one can takeΛ = 0, but for noisy data takingΛ = 0
produces a highly unbounded and oscillatory solution forr′(t) andP (t), as shown in
Figures 4.5(c) and 4.5(d). For noisy data, reference [59] suggests thea priori choice
Λ =T
N
N∑
j=1
(rexact,j − rj)2, (4.34)
whererexact,j = rexact(tj) represents an analytical solutionr, given by (4.27), at time
tj for j = 1, N . Forp = 1% noise, we haveΛ ≈ 8 × 10−3. The numerical results for
r′(t) andP (t) obtained using the smoothing spline regularisation technique with this
choice ofΛ oversmoothes the exact solution, as shown in Figure 4.6 in dash-dot line
(− · −). However, the Morozov’s discrepancy principle [40] based on thea posteriori
choice ofΛ such that
T
N
N∑
j=1
(rΛ(tj)− rj)2 ≈ T
N
N∑
j=1
(rexact,j − rj)2, (4.35)
producesΛdis = 7 × 10−5 which yields more accurate approximations, as also shown
Chapter 4. 97
0 0.2 0.4 0.6 0.8 11
2
3
4
5
6
7
8
9
t
r′(t
)
(a)
0 0.2 0.4 0.6 0.8 10.5
1
1.5
2
2.5
3
t
P(t
)
(b)
Figure 4.6: The analytical (—–) and numerical results of (a)r′(t) and (b)P (t) obtainedusing the smoothing spline regularisation withΛdis = 7×10−5 () andΛ = 8×10−3
(− · −), as defined in (4.34), forp = 1% noise, for Example 2.
in Figure 4.6 in circle line( ).Finally, although not illustrated, it is reported that for higher amounts of noise the
numerical reconstructions are less stable.
4.6 Conclusions
The inverse problem of finding a time-dependent blood perfusion coefficient for the
bioheat equation with nonlocal boundary conditions and mass/energy specification has
been investigated. The inverse problem has been transformed to an inverse heat source
problem with an unknown term present in the integral over-determination condition.
The numerical discretisation was based on the BEM together with the Tikhonov reg-
ularisation and the GCV for the choice of the regularisationparameter. We have also
applied the smoothing spline technique for differentiating a noisy function witha pri-
ori anda posteriorichoices of the regularisation parameters. For a couple of typical
benchmark test examples, accurate and stable numerical solutions have been obtained.
Chapter 5. 98
Chapter 5
Determination of a Time-dependent
Heat Source with a Dynamic
Boundary Condition
5.1 Introduction
In the Chapters 3 and 4, we have considered inverse time-dependent source problems
for the heat equation with various types of conditions such as integral, local or nonlo-
cal. In the present chapter, we consider yet another reconstruction of a time-dependent
heat source with the integral over-determination measurement of the thermal energy
of the system and a new dynamic-type boundary condition. This model can be used
in heat transfer and diffusion processes with a time-dependent source parameter to be
determined. Also, in acoustic scattering or damage corrosion the new dynamic-type
boundary condition (5.4) below is also known as a generalised impedance boundary
condition, [4–7].
The well-posedness of the inverse problem studied in this chapter was established
in [19], and we aim to obtain the numerical solution by using the BEM together with a
regularisation method.
99
Chapter 5. 100
This chapter is organised as follows. In Section 5.2, the mathematical formula-
tion of the inverse problem is described. The numerical discretisation of the problem
based on the BEM is described in Section 5.3. Section 5.4 discusses numerical results
obtained for three of benchmark test examples and emphasises the importance of em-
ploying the Tikhonov regularisation with the choice of regularisation parameter based
on either the GCV criterion or the discrepancy principle, inorder to achieve a stable
numerical solution. Finally, Section 5.5 presents the conclusions of the chapter.
5.2 Mathematical formulation
Consider the following initial-boundary value problem of finding the time-depending
heat sourcer(t) ∈ C([0, T ]) and the temperatureu(x, t) ∈ C2,1(DT ) ∩ C1,0(DT )
which satisfy the heat equation:
ut = uxx + r(t)f(x, t), (x, t) ∈ DT , (5.1)
whereL = 1 in the definition ofDT in (1.1), subject to the initial condition (1.7),
namely
u(x, 0) = u0(x), x ∈ [0, 1], (5.2)
and the boundary conditions
u(0, t) = 0, t ∈ (0, T ], (5.3)
auxx(1, t) + αux(1, t) + bu(1, t) = 0, t ∈ (0, T ], (5.4)
wheref andu0 are given functions anda, b, α are given numbers not simultaneously
equal to zero. The well-posedness of this direct problem wasestablished in [34].
Taking into account the equation (5.1) atx = 1, the boundary condition (5.4)
Chapter 5. 101
becomes
aut(1, t) + αux(1, t) + bu(1, t) = ar(t)f(1, t), t ∈ (0, T ]. (5.5)
In order to add further physics to the problem, we mention that the boundary condition
(5.5) is observed in the process of cooling of a thin solid barone end of which is placed
in contact with a fluid [36]. Another possible application ofsuch type of boundary
condition is announced in [9, p.79], as this boundary condition represents a boundary
reaction in diffusion of a chemical. We finally mention that we have also previously
encountered the dynamic boundary condition (5.5) when modelling a transient flow
pump experiment in a porous medium [39].
When the functionr(t) for t ∈ [0, T ] is unknown, the inverse problem formulates
as a problem of finding a pair of functions(r(t), u(x, t)) which satisfy the equation
(5.1), initial condition (5.2), the boundary conditions (5.3) and (5.4) (or (5.5)), and the
energy/mass overdetermination measurement
∫ 1
0
u(x, t) dx = E(t), t ∈ [0, T ]. (5.6)
This overdetermination condition is encountered in modelling applications related to
particle diffusion in turbulent plasma, as well as in heat conduction problems in which
the law of variationE(t) of the total energy of heat in a rod is given, [22].
If we let u(x, t) represent the temperature distribution, then the above-mentioned
inverse problem can be regarded as a source control problem.The source control
parameterr(t) needs to be determined from the measurement of the thermal energy
E(t).
DenoteΦ4n0[0, 1] := φ(x) ∈ C4[0, 1];φ(0) = φ′′(0) = 0, φ(1) = φ′(1) =
φ′′(1) = φ′′′(1) = 0,∫ 1
0φ(x)yn0
(x)dx = 0, where(yn)n≥0 denote the eigenfunctions
Chapter 5. 102
of the spectral problem
y′′(x) + µy(x) = 0, x ∈ [0, 1],
y(0) = 0,
(aµ− b)y(1) = αy′(1),
(5.7)
where its eigenvalueµn and eigenfunctionsyn(x), for n = 0, 1, 2, . . . , have the fol-
lowing asymptotic behavior:
√µn = πn +O
(
1
n
)
, yn(x) = sin(πnx) +O
(
1
n
)
,
for sufficiently largen. The following theorem proved in [19] established the existence
of a unique solution of the inverse problem (5.1)–(5.3), (5.5) and (5.6).
Theorem 5.2.1 Letaα > 0 and assume that the following conditions are satisfied:
(A1) u0(x) ∈ Φ4n0[0, 1];
(A2) E(t) ∈ C1[0, T ]; E(0) =∫ 1
0u0(x)dx;
(A3) f(x, t) ∈ C(DT ); f(·, t) ∈ Φ4n0[0, 1], ∀t ∈ [0, T ]; and
∫ 1
0f(x, t)dx 6= 0, ∀t ∈
[0, T ];
Then the inverse problem(5.1)–(5.3), (5.5)and(5.6)has a unique solution(r(t), u(x, t)) ∈C[0, T ]× (C2,1(DT ) ∩ C2,0(DT )). Moreover,u(x, t) ∈ C2,1(DT ).
Although the inverse problem is uniquely solvable, it is still ill-posed since small errors
in the input data (5.6) cause large errors in the input sourcer(t).
In the next section we will describes the discretisation of the inverse problem using
the BEM, whilst Section 5.4 will discuss the regularisationof the numerical solution.
Chapter 5. 103
5.3 Boundary element method (BEM)
In this section, we explain the numerical procedure for discretising the inverse problem
(5.1)–(5.3), (5.5) and (5.6) by using the BEM. As introducedin Section 1.3, the use
of BEM recasts the heat equation (5.1) in the boundary integral form (3.5). Using the
same discretisation as described in Chapters 2–4, and applying the boundary condition
(5.3), the boundary integral equation (3.5) becomes
η(x)u(x, t) =
N∑
j=1
[A0j(x, t)q0j + ALj(x, t)qlj −BLj(x, t)hLj ]
+
N0∑
k=1
Ck(x, t)u0,k +N∑
j=1
Dj(x, t)rj , (x, t) ∈ [0, 1]× (0, T ]. (5.8)
On applying the BEM and the boundary condition (5.3), we obtain the system of2N
linear equations,
A0q¯0
+ ALq¯L
− BLh¯L
+ Cu¯0
+Dr¯= 0
¯. (5.9)
In order to apply the boundary condition (5.5) we need to approximate the time-
derivativeut(1, t) by using finite differences. For this, we use theO(h2) finite differ-
ence formulae
ut(1, t1) =u(1, t2)/3 + u(1, t1)− 4u0(1)/3
h,
ut(1, t2) =5u(1, t2)/3− 3u(1, t1) + 4u0(1)/3
h,
ut(1, ti) =3u(1, ti)/2− 2u(1, ti−1) + u(1, ti−2)/2
h, i = 3, N,
where the step sizeh = T/N and ti = ti−1+ti2
as defined in Chapter 1. Applying
the expressions above into the boundary condition (5.5) yields the linear system ofN
Chapter 5. 104
equations as follows:
a
hu(1, t1) +
a
3hu(1, t2) + bu(1, t1) = ar(t1)f(1, t1) +
4a
3hu0(1)− αux(1, t1),
−3a
hu(1, t1) +
5a
3hu(1, t2) + bu(1, t2) = ar(t2)f(1, t2)−
4a
3hu0(1)− αux(1, t2),
a
2hu(1, ti−2)−
2a
hu(1, ti−1) +
3a
2hu(1, ti) + bu(1, ti) = ar(ti)f(1, ti)− αux(1, ti),
for i = 3, N . This system can be rewritten as
Sh¯L
= Fr¯+ u
¯0− αq
¯L, (5.10)
whereF = diag(af(1, t1), . . . , af(1, tN)), and
S =
a/h + b a/3h 0 .
−3a/h 5a/3h+ b 0 .
a/2h −2a/h 3a/2h+ b .
. . . .
0 a/2h −2a/h 3a/2h+ b
N×N
, u¯0
=
4au0(1)/3h
−4au0(1)/3h
0
.
0
N
.
Assumingα 6= 0, eliminating q¯L
can be done by applying the derived matrix form of
(5.10), i.e.
q¯L
=1
α(Fr
¯+ u
¯0− Sh
¯L) , (5.11)
into (5.9), this gives (5.9) and (5.10) results as
h¯L
q¯0
=
[(
1
αALS +BL
)
∣
∣
∣−A0
]−1(1
αALFr
¯+
1
αALu
¯0+ Cu
¯0+Dr
¯
)
, (5.12)
where the invertible matrix[
(
1αALS +BL
)
∣
∣
∣−A0
]
is a 2N × 2N matrix formed
with the 2N × N block matrices(
1αALS +BL
)
and−A0 separated by the vertical
line.
Next, we collocate the over-determination condition (5.6)by using the midpoint
Chapter 5. 105
numerical integration approximation as same as in (3.11). Then the expression (5.8) at
(xk, ti), can be rewritten as
1
N0
N0∑
k=1
[
A(1)0,kq
¯0+ A
(1)L,kq
¯L− B
(1)L,kh¯L
+ C(1)k u
¯0+D
(1)k r
¯
]
= E¯, (5.13)
where
A(1)0,k =
[
A0j(xk, ti)]
N×N, A
(1)L,k =
[
ALj(xk, ti)]
N×N, B
(1)L,k =
[
BLj(xk, ti)]
N×N,
C(1)k =
[
Cl(xk, ti)]
N×N0
, D(1)k =
[
Dj(xk, ti)]
N×N, k, l = 1, N0, i, j = 1, N.
When the boundary condition (5.3) is applied, the expression (5.13) can be, via (5.11),
rewritten as
1
N0
N0∑
k=1
[
A(1)0,kq
¯0+
1
αA
(1)L,k (Fr
¯+ u
¯0− Sh
¯L)− B
(1)L,kh¯L
+ C(1)k u
¯0+D
(1)k r
¯
]
= E¯. (5.14)
Finally, eliminating q¯0
and h¯L
between (5.12) and (5.14), the unknown discretised
source r¯
can be found by solving theN ×N linear system of equations
Xr¯= y
¯, (5.15)
where
X =1
N0
N0∑
k=1
[(
1
αA
(1)L,kS +B
(1)L,k
)
∣
∣
∣−A(1)
0,k
] [(
1
αALS +BL
)
∣
∣
∣−A0
]−1
×(
1
αALF +D
)
−(
1
αA
(1)L,kF +D
(1)k
)
,
y¯=
1
N0
N0∑
k=1
C(1)k u
¯0+
1
αA
(1)L,ku¯0
−[(
1
αA
(1)L,kS +B
(1)L,k
)
∣
∣
∣−A(1)
0,k
]
×[(
1
αALS +BL
)
∣
∣
∣−A0
]−1(
Cu¯0
+1
αALu
¯0
)
− E¯.
Chapter 5. 106
As we have mentioned previously the inverse problem under investigation is ill-
posed, and consequently, the system of equation (5.15) is ill-conditioned. In the next
section, we will discuss the regularisation of the numerical solution together with
choices of regularisation parameter based on either the GCVcriterion or the discrep-
ancy principle.
5.4 Numerical examples and discussion
This section presents three benchmark test examples with smooth and non-smooth con-
tinuous source functions in order to test the accuracy of theBEM numerical procedure
introduced earlier in Section 5.3. In order to illustrate the accuracy of the numerical
results, the RMSE defined in (2.49) is also used here.
5.4.1 Example 1
In the first example, we consider the case of smooth continuous unknown source func-
tion, given by the analytical solution
u(x, t) = x2et, r(t) = et, (5.16)
for the inverse problem (5.1)–(5.3), (5.5) and (5.6) with the input dataT = 1, a = α =
1, b = −4,
u(x, 0) = u0(x) = x2, f(x, t) = x2 − 2, (5.17)
The direct problem (5.1)–(5.3) and (5.5), whenr(t) = et is known, is considered first
with N = N0 ∈ 20, 40, 80 obtained by (5.11), (5.12) and (5.13), and the RMSE
results are shown in Table 5.1. From this table it can be concluded that the BEM
numerical solution is convergent to the corresponding exact values
u(1, t) = et, ux(0, t) = 0, ux(1, t) = 2et, E(t) = et/3, (5.18)
Chapter 5. 107
for t ∈ [0, 1], as the number of boundary elements increases.
Table 5.1: The RMSE foru(1, t), ux(0, t), ux(1, t) andE(t) obtained using the BEMfor the direct problem withN = N0 ∈ 20, 40, 80, for Example 1.
N = N0RMSE
u(1, t) ux(0, t) ux(1, t) E(t)20 6.43E-3 2.79E-3 8.85E-3 2.65E-340 2.20E-3 9.68E-4 2.98E-3 9.07E-480 7.46E-4 3.32E-4 1.00E-3 3.08E-4
Next, we consider the inverse problem (5.1)–(5.3), (5.5) and (5.6) and we use the
BEM with N = N0 = 40 for solving the resulting system of equations (5.15). Figure
5.1 displays the analytical and numerical results ofr(t), u(1, t), ux(0, t), andux(1, t)
and very good agreement can be observed.
In practice, the contamination of measured data by unplanned error is unavoidable.
Thus we add noise to the input energy dataE(t) in (5.6) in order to test the stability of
the solution. Here, the perturbed input data E¯ǫ is defined as same as described in (3.14),
with the standard deviationσ =ep
3andp ∈ 1, 3, 5%. This perturbation means that
the known right-hand side vector y¯
of the linear system (5.15) is contaminated with
noise, denoted as y¯ǫ. Then, when noise is present, we have to solve the following
system of linear equations instead of (5.15):
Xr¯= y
¯ǫ. (5.19)
Initially, we have tried to solve the above disturbed system(5.19) withp = 1% noise
in the input data (5.6) by using the straightforward inversion of (5.19), i.e. r¯= X−1y
¯ǫ,
illustrated in Figure 5.2. From this figure it can be seen thatthe numerical solutions for
r(t), ux(0, t) andux(1, t) shown by the dash-dot line (− · −) are unstable. However,
the result foru(1, t) seems to remain stable.
To overcome this instability, we employ the Tikhonov regularisation method as
described in (1.34) with a second-order differential regularisation matrix in (4.20). As
happened previously with some of our investigations in Chapters 2 and 3, we report that
Chapter 5. 108
0 0.2 0.4 0.6 0.8 11
1.5
2
2.5
3
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 11
1.5
2
2.5
3
t
u(1
,t)
(b)
0 0.2 0.4 0.6 0.8 1−3
−2.5
−2
−1.5
−1
−0.5
0
0.5
1x 10
−3
t
ux(0
,t)
(c)
0 0.2 0.4 0.6 0.8 12
2.5
3
3.5
4
4.5
5
5.5
t
ux(1
,t)
(d)
Figure 5.1: The analytical (—–) and numerical results(− · −) of (a) r(t), (b) u(1, t),(c) ux(0, t), and (d)ux(1, t) for exact data, for Example 1.
the second-order Tikhonov regularisation has produced more accurate results than the
zeroth- or first-order regularisation and therefore, only the numerical results obtained
using the former regularisation are illustrated in this section.
A popular method for choosing the regularisation parameteris the GCV criterion
which is based on the minimisation technique, as we have detailed in Section 1.6. For
p = 1% noise, this minimisation yields the minimum point of the GCVfunction (1.35)
occurring atλGCV =4.3E-6. Then the numerical results obtained using the Tikhonov
regularisation with this value ofλGCV , illustrated by circles( ) in Figure 5.2, show
that accurate and stable numerical solutions are achieved.
Next, we increase top = 3% and5% the percentage of noise. Figure 5.3 presents
Chapter 5. 109
0 0.2 0.4 0.6 0.8 1−4
−2
0
2
4
6
8
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 11
1.5
2
2.5
3
t
u(1
,t)
(b)
0 0.2 0.4 0.6 0.8 1−0.4
−0.3
−0.2
−0.1
0
0.1
0.2
0.3
0.4
t
ux(0
,t)
(c)
0 0.2 0.4 0.6 0.8 11.5
2
2.5
3
3.5
4
4.5
5
5.5
t
ux(1
,t)
(d)
Figure 5.2: The analytical (—–) and numerical results of (a)r(t), (b) u(1, t), (c)ux(0, t), and (d)ux(1, t) obtained using the straightforward inversion(− · −) withno regularisation, and the second-order Tikhonov regularisation( ) with the regu-larisation parameterλ=4.3E-6 suggested by the GCV method, forp = 1% noise, forExample 1.
the analytical and numerical results obtained using the second-order Tikhonov reg-
ularisation with the regularisation parameter suggested by the GCV method, namely
λGCV =7.4E-6 forp = 3%, andλGCV =2.7E-5 forp = 5%. From this figure one can ob-
serve that stable and accurate results forr(t), u(1, t), ux(0, t) andux(1, t) with p = 3%
noise are attained, whereas the numerical results forp = 5% noisy input are rather in-
accurate, but they remain stable. For completeness, the RMSE errors are displayed in
Table 5.2.
Chapter 5. 110
0 0.2 0.4 0.6 0.8 10.5
1
1.5
2
2.5
3
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 11
1.5
2
2.5
3
t
u(1
,t)
(b)
0 0.2 0.4 0.6 0.8 1−0.4
−0.2
0
0.2
t
ux(0
,t)
(c)
0 0.2 0.4 0.6 0.8 12
2.5
3
3.5
4
4.5
5
5.5
t
ux(1
,t)
(d)
Figure 5.3: The analytical (—–) and numerical results of (a)r(t), (b) u(1, t), (c)ux(0, t), and (d)ux(1, t) obtained using the second-order Tikhonov regularisation withthe regularisation parameter suggested by the GCV method, for p = 3% (· · ·) andp = 5% (−−−), for Example 1.
5.4.2 Example 2
The previous example possessed an analytical solution being explicitly available; how-
ever the source functionf(x, t) chosen did not satisfy the condition in (A3) of Theorem
5.2.1 thatf ∈ Φ4n0[0, 1]. Therefore, in this subsection we aim to construct an example
for which the conditions of existence and uniqueness of solution of Theorem 5.2.1 are
satisfied. We chooseT = 1, a = α = 1, b = 0 andu0(x) = 0.
In the casea = α = 1, b = 0 the problem (5.7) has the eigenvaluesµn = ν2n,
whereνn are the positive roots of the transcendental equationν sin(ν) = cos(ν). The
Chapter 5. 111
Table 5.2: The regularisation parametersλ and the RMSE forr(t), u(1, t), ux(0, t) andux(1, t), obtained using the BEM withN = N0 = 40 combined with the second-orderTikhonov regularisation forp ∈ 0, 1, 3, 5% noise, for Example 1.
pparameter RMSE
λ r(t) u(1, t) ux(0, t) ux(1, t)0 0 4.16E-3 2.47E-4 1.20E-3 8.85E-4
1% 0 2.70 1.72E-2 1.12E-1 2.64E-11% λGCV =4.3E-6 1.73E-2 2.57E-3 8.92E-3 5.47E-33% 0 5.21 4.13E-2 3.51E-1 5.02E-13% λGCV =7.4E-6 3.32E-2 9.73E-3 1.97E-2 2.25E-25% 0 4.74 5.51E-2 4.64E-1 4.54E-15% λGCV =2.7E-5 1.95E-1 4.63E-2 1.29E-1 9.79E-2
corresponding eigenfunctions areyn(x) = sin(νnx). The first eigenvalue is given by
ν0 =√µ0 = 0.860333. Then choosing
f(x, t) = x3(1− x)4(β1x+ β2), (5.20)
we can determine the constantsβ1 andβ2 such thatf ∈ Φ40[0, 1] (choosingn0 = 0 for
simplicity), as required by the condition (A3) of Theorem 5.2.1, i.e.f(x, t) ∈ Φ4n0[0, 1],
∀t ∈ [0, t]. This imposes
0 =
∫ 1
0
f(x, t) sin(ν0x) dx =
∫ 1
0
x3(1− x)4(β1x+ β2) sin(ν0x) dx.
After some calculus, choosingβ2 = −1 it follows thatβ1 ≈ 2.011. With these values
of β1 andβ2 we also satisfy that∫ 1
0f(x, t) dx = −0.00037 is non-zero, as required by
condition (A3). We aim to retrieve a non-smooth source function given by
r(t) =
∣
∣
∣
∣
t− 1
2
∣
∣
∣
∣
, t ∈ [0, t]. (5.21)
In this case, the analytical solution of the direct problem for the temperatureu(x, t)
is not available. Thus the energyE(t) is not available either. In such a situation, we
simulate the data (5.6) numerically by solving first the direct problem (5.1)–(5.3) and
Chapter 5. 112
(5.5) with r known and given by (5.21). The numerical solutions foru(1, t), ux(0, t),
ux(1, t) andE(t) obtained using the BEM withN = N0 ∈ 20, 40, 80 are shown
in Figure 5.4. From this figure it can be seen that convergent numerical solutions are
obtained.
0 0.2 0.4 0.6 0.8 1−0.5
0
0.5
1
1.5
2
2.5
3
3.5x 10
−6
t
u(1
,t)
(a)
0 0.2 0.4 0.6 0.8 1−1.8
−1.6
−1.4
−1.2
−1
−0.8
−0.6
−0.4
−0.2
0x 10
−4
t
ux(0
,t)
(b)
0 0.2 0.4 0.6 0.8 1−4
−3
−2
−1
0
1x 10
−5
t
ux(1
,t)
(c)
0 0.2 0.4 0.6 0.8 1−1.4
−1.2
−1
−0.8
−0.6
−0.4
−0.2
0x 10
−5
t
E(t
)
(d)
Figure 5.4: The numerical results of (a)u(1, t), (b) ux(0, t), (c) ux(1, t), and (d)E(t)obtained by solving the direct problem withN = N0 ∈ 20(), 40(···), 80(−−−),for Example 2.
To investigate the inverse problem (5.1)–(5.3), (5.5) and (5.6), we use the numerical
results forE(t) in Figure 5.4(d) obtained using the BEM withN = N0 = 40, as the
input data (5.6). In order to avoid committing an inverse crime we keepN = 40, but
we use a differentN0, sayN0 = 30, than 40 which was used in the direct problem
simulation. Figure 5.5 shows the numerical results obtained without regularisation,
Chapter 5. 113
i.e. λ = 0, for p = 0 (exact) andp = 1% (noisy) data. Remark that from Figure 5.4(d),
the standard deviation for generating noise is given byσ = 1.2 × 10−5p. From Figure
5.5 it can be seen that, for exact data, the straightforward inversion of (5.15) produces
very accurate results. However, when noise is introduced into the measured data E¯ǫ,
here we are solving the linear system of equationsXr¯= y
¯ǫ instead ofXr
¯= y
¯, the
numerical retrievals of especiallyr(t) andux(1, t) become highly oscillatory unstable.
In order to retrieve the stability, as in Example 1, the second-order Tikhonov reg-
ularisation with the GCV criterion are employed and the numerically obtained results
are shown in Figure 5.6. The numerical results from the direct problem presented in
Figures 5.7(a)–5.7(c) are used to compare in Figures 5.6(b)–5.6(d) the numerical re-
sults foru(1, t), ux(0, t), andux(1, t), respectively, of the inverse problem. Whereas
the numerical solution forr(t) of the inverse problem is compared with the analytical
solution (5.21) in Figure 5.6(a). From Figure 5.6 it can be seen that stable and accurate
numerical solutions are obtained. For completeness, the RMSE errors and the GCV
values forλ are displayed in Table 5.3.
Table 5.3: The regularisation parametersλ and the RMSE forr(t), u(1, t), ux(0, t)andux(1, t), obtained using the BEM withN = 40 andN0 = 30 combined with thesecond-order Tikhonov regularisation forp ∈ 0, 1, 3, 5% noise, for Example 2.
p(%)parameter RMSE
λ r(t) u(1, t) ux(0, t) ux(1, t)0 0 2.90E-4 2.12E-10 2.94E-8 1.07E-8
1% 0 9.98E-2 2.71E-8 9.55E-6 3.55E-61% λGCV =3.2E-16 5.47E-3 1.62E-8 1.28E-6 4.12E-73% 0 3.63E-1 1.05E-7 3.39E-5 1.27E-53% λGCV =1.1E-15 1.37E-2 5.96E-8 3.56E-6 9.37E-75% 0 5.03E-1 1.70E-7 4.85E-5 1.80E-55% λGCV =9.0E-16 2.17E-2 9.88E-8 5.68E-6 1.21E-6
If one would like to make a fair comparison between the accuracy of the numerical
results obtained for Examples 1 and 2, the RMSE values presented in Tables 5.2 and
5.3 should be divided by the maximum absolute values of the corresponding quantities
involved. For example, if we divide the columns of RMSE values for r(t) in Tables
Chapter 5. 114
0 0.2 0.4 0.6 0.8 1−0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 1
0
0.5
1
1.5
2
2.5
3
3.5x 10
−6
t
u(1
,t)
(b)
0 0.2 0.4 0.6 0.8 1−2
−1.5
−1
−0.5
0x 10
−4
t
ux(0
,t)
(c)
0 0.2 0.4 0.6 0.8 1−4
−3
−2
−1
0
1
2x 10
−5
t
ux(1
,t)
(d)
Figure 5.5: The analytical solution (5.21) and the direct problem numerical solutionfrom Figures 5.7(a)–5.7(c) (—–) and numerical results of (a) r(t), (b) u(1, t), (c)ux(0, t), and (d)ux(1, t), with no regularisation, for exact data( ) and noisy datap = 1% (− · −), for Example 2.
5.2 and 5.3 bye (maximum value ofr(t) in (5.16)) and0.5 (maximum value ofr(t) in
(5.21)), respectively, then the relative errors forr(t) in Example 1 are actually lower
than those in Example 2, as expected from the regularity of these solution.
Finally, although not illustrated, it is reported that for both Examples 1 and 2 we
have experienced with other values ofλ close to the optimal ones but there was not
much difference obtained in comparison with the numerical results of Figures 5.2, 5.3
and 5.6. This confirms that the GCV criterion performs well inchoosing a suitable
regularisation parameter for obtaining a stable and accurate numerical solution.
Chapter 5. 115
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 1
0
0.5
1
1.5
2
2.5
3
3.5x 10
−6
t
u(1
,t)
(b)
0 0.2 0.4 0.6 0.8 1−2
−1.5
−1
−0.5
0x 10
−4
t
ux(0
,t)
(c)
0 0.2 0.4 0.6 0.8 1
−3
−2
−1
0
1
x 10−5
t
ux(1
,t)
(d)
Figure 5.6: The analytical solution (5.21) and the direct problem numerical solu-tions from Figures 5.7(a)–5.7(c) (—–), and the numerical results of (a)r(t), (b)u(1, t), (c) ux(0, t), and (d)ux(1, t) obtained using the second-order Tikhonov reg-ularisation with the regularisation parameters suggestedby GCV method, forp ∈1( ), 3(· · · ), 5(−−−)% noise, for Example 2.
5.4.3 Example 3
In previous examples, we have use the BEM together with the second-order Tikhonov
regularisation with the value of the regularisation parameter suggested by the GCV
method, on both cases for identification of the smooth and non-smooth source func-
tions in Example 1 and 2, respectively. In this example, we consider yet another case
Chapter 5. 116
example of non-smooth source function given by
r(t) =
∣
∣
∣
∣
t− 1
2
∣
∣
∣
∣
,
with the input dataT = 1, a = α = 1, b = 0,
u0(x) = 0, f(x, t) = 1. (5.22)
We remark that the condition in(A3) of Theorem 5.2.1 does not hold.
In this case, as with Example 2 that the analytical solution of the direct problem
for the temperatureu(x, t) is not available, thus the energyE(t) is not available either.
Therefore as we have done in previous example, the direct problem has been solved
first with the BEM andN = N0 ∈ 20, 40, 80, as illustrated in Figure 5.7. From
this figure it can be seen that rapidly convergent numerical solutions are obtained. The
numerical results obtained using the BEM withN = N0 = 40 are kept as an input data
for E(t) and the reference (analytical) solution foru(1, t), ux(0, t) andux(1, t).
In what follows, we present numerical results obtained withN = 40 andN0 = 30
instead of 40, in order to avoid committing an inverse crime.The energy dataE(t)
obtained from the numerical result in Figure 5.7(d) is perturbed, as described in (3.14)
with the standard deviationσ = 0.15p. Figure 5.8 shows that the numerical results
obtained by the straightforward inversion of (5.15), i.e. r¯= X−1y
¯, without regularisa-
tion, for p = 0 (exact) andp = 1% (noisy) data. For exact data, the same very good
agreement between the analytical and numerical solutions is recorded. Whereas for
the noisy data, the numerical solutions forr(t) andux(1, t) becomes highly oscilla-
tory and unbounded. This is somewhat to be expected since theinverse problem under
investigation is ill-posed.
We retrieve the stability, as previous, by the second-orderTikhonov regularisation
with the proper choice of the regularisation parameterλ. Firstly, we have used the rig-
orous discrepancy principle, as introduced in Section 1.6.3, for p ∈ 1, 3, 5% which
yieldsλdis ∈ 5.7E-8, 2.0E-6, 7.0E-6, respectively. The results are obtained as shown
Chapter 5. 117
0 0.2 0.4 0.6 0.8 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
t
u(1
,t)
(a)
0 0.2 0.4 0.6 0.8 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
t
ux(0
,t)
(b)
0 0.2 0.4 0.6 0.8 10
0.05
0.1
0.15
0.2
0.25
t
ux(1
,t)
(c)
0 0.2 0.4 0.6 0.8 10
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
t
E(t
)
(d)
Figure 5.7: The numerical results of (a)u(1, t), (b) ux(0, t), (c) ux(1, t), and (d)E(t)obtained by solving the direct problem withN = N0 ∈ 20 (− ·−), 40 (· · ·), 80 (−−−), for Example 3.
in Figure 5.9 and Table 5.4. From Figure 5.9 it can be seen thatstable and accurate
numerical solutions are obtained. Table 5.4 also shows thatthe values of the regular-
isation parameterλdis andλGCV given by the discrepancy principle and the GCV are
similar and not very different from the optimal valueλopt given by the minimum of the
RMSE forr(t). It can also be seen that the GCV produces slightly better predictions
than the discrepancy principle.
5.5 Conclusions
The inverse problem of finding the time-dependent heat source together with the tem-
perature in the heat equation, under a non-classical dynamic boundary condition and
Chapter 5. 118
0 0.2 0.4 0.6 0.8 1−1
−0.5
0
0.5
1
1.5
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 10
0.05
0.1
0.15
0.2
0.25
0.3
t
u(1
,t)
(b)
0 0.2 0.4 0.6 0.8 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
t
ux(0
,t)
(c)
0 0.2 0.4 0.6 0.8 1−0.05
0
0.05
0.1
0.15
0.2
0.25
t
ux(1
,t)
(d)
Figure 5.8: The analytical solution (5.4.3) and the direct problem numerical solutionfrom Figures 5.7(a)–5.7(c) (—–), and numerical results of (a) r(t), (b) u(1, t), (c)ux(0, t), and (d)ux(1, t), with no regularisation, for exact data( ) and noisy datap = 1% (− · −), for Example 3.
an integral over-determination condition has been investigated. A numerical method
based on the BEM combined with the second-order Tikhonov regularisation has been
proposed together with the use of either the GCV criterion orthe discrepancy principle
for the selection of the regularisation parameter. The retrieved numerical results were
found to be accurate and stable on both smooth and non-smoothcontinuous examples.
As for the experimental validation of the proposed inverse mathematical model in
terms of bias and inverting real noisy data we defer this challenging task to possible
future work. We only remark that unlike certain applications, e.g. some significant
mismatch has been reported in [2, 35, 52] between experimental data of electromag-
netic waves propagating in a non-attenuating medium and data produced by idealised
computational simulations, in inverse heat conduction themathematical models have
Chapter 5. 119
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 10
0.05
0.1
0.15
0.2
0.25
0.3
t
u(1
,t)
(b)
0 0.2 0.4 0.6 0.8 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
t
ux(0
,t)
(c)
0 0.2 0.4 0.6 0.8 10
0.05
0.1
0.15
0.2
0.25
t
ux(1
,t)
(d)
Figure 5.9: The analytical solution (5.4.3) and the direct problem numerical solutionsfrom Figures 5.7(a)–5.7(c) (—–), and the numerical resultsof (a) r(t), (b) u(1, t), (c)ux(0, t), and (d)ux(1, t) obtained using the second-order Tikhonov regularisation withthe regularisation parameters suggested by the discrepancy method, forp ∈ 1(− ·−), 3(· · · ), 5(−−−)% noise, for Example 3.
been shown to perform much better in industrial applications with actual real measured
data, [13].
So far in previous chapters, the identification of a single time-dependent heat source
function has been the main aim. In the next two chapters, we will consider the more
challenging cases of simultaneous determination of the components of an additive or
multiplicative space- and time-dependent heat source function.
Chapter 5. 120
Table 5.4: The regularisation parametersλ given by the discrepancy principle andby the GCV, and the RMSE forr(t), u(1, t), ux(0, t) and ux(1, t), obtained usingthe BEM withN = 40 andN0 = 30 combined with the second-order Tikhonovregularisation forp ∈ 0, 1, 3, 5% noise. The optimal regularisation parameter givenby the minimum of RMSE ofr(t) is also included, for Example 3.
pparameter RMSE
λ r(t) u(1, t) ux(0, t) ux(1, t)0 0 2.46E-4 8.06E-6 1.41E-5 2.27E-5
1% 0 5.13E-1 3.26E-3 1.20E-2 4.19E-21% λdis=5.7E-8 1.11E-2 9.25E-4 2.89E-3 5.82E-41% λGCV =5.7E-9 1.19E-2 8.26E-4 2.76E-3 5.96E-41% λopt=2.2E-8 9.85E-3 8.06E-4 2.46E-3 5.65E-43% 0 1.53 9.86E-3 3.71E-2 1.23E-13% λdis=2.0E-6 3.13E-2 3.67E-3 9.99E-3 2.85E-33% λGCV =3.3E-7 3.01E-2 2.53E-3 8.08E-3 1.81E-33% λopt=9.0E-7 2.75E-2 2.77E-3 8.05E-3 2.07E-35% 0 1.30 1.31E-2 6.26E-2 1.03E-15% λdis=7.0E-6 7.86E-2 8.95E-3 2.68E-2 5.25E-35% λGCV =7.9E-7 4.57E-2 6.80E-3 1.72E-2 4.43E-35% λopt=9.1E-8 4.15E-2 6.68E-3 1.59E-2 4.61E-3
Chapter 6
Determination of an Additive Space-
and Time-dependent Heat Source
6.1 Introduction
Inverse source problems for the heat equation have recentlyattracted considerable in-
terest, see [1, 18, 24, 61, 64, 67] to name just a few. These studies have sought a
coefficient source function depending on either space- or time-dependent variables us-
ing various techniques. In Chapters 2–5 we have focused on the inverse problem of
finding the time-dependent coefficient source function and the temperature for the heat
equation. In this chapter we extend our study to determine inverse heat source func-
tions depending on both space and time, but which are additively separated into two
unknown coefficient source functions, namely, one component dependent on space and
another component dependent on time. The measurement/overspecified conditions are
one temperature measurement, as a function of time, at one specified interior location
and a time-average temperature throughout the space solution domain.
Since the governing partial differential equation is the linear heat equation with
constant coefficients, the preferred method of discretisation is the BEM. Even though
the inverse heat source problem is uniquely solvable, it is still ill-posed since small
121
Chapter 6. 122
errors which inherently occur in any practical measurementcause largely oscillating
solutions. To overcome this instability, in this chapter regularisation such as the TSVD
or the Tikhonov regularisation method are employed. Moreover, theL-curve method,
the GCV criterion, or the discrepancy principle are employed for the selection of the
regularisation parameter. Additionally in the case of two different regularisation pa-
rameters are considered, theL-surface criterion, [3], is utilised.
This chapter is organised as follows. In Section 6.2 the mathematical inverse for-
mulation is described and the numerical discretisation of the problem using the BEM
is presented in Section 6.3. The TSVD and the Tikhonov regularisation are described
in Section 6.4, as procedures for overcoming the instability of the solution. Finally,
Sections 6.5 and 6.6 discuss the numerical results and highlight the conclusions of this
research.
6.2 Mathematical formulation
Consider the inverse problem of finding the time-dependent heat sourcer(t), the space-
dependent heat sources(x) and the temperatureu(x, t) satisfying the heat conduction
equation
ut = uxx + r(t)f(x, t) + s(x)g(x, t) + h(x, t), (x, t) ∈ DT , (6.1)
subject to the initial condition (1.7), namely
u(x, 0) = u0(x), x ∈ [0, L], (6.2)
the Dirichlet boundary conditions
u(0, t) = µ0(t), u(L, t) = µL(t), t ∈ [0, T ], (6.3)
Chapter 6. 123
and the over-determination conditions
u (X0, t) = χ(t), t ∈ [0, T ], (6.4)∫ T
0
u(x, t) dt = ψ(x), x ∈ [0, L], (6.5)
s(X0) = S0, (6.6)
whenf , g, h, u0, µ0, µL, χ, ψ are given functions,X0 is a given sensor location within
the interval(0, L), andS0 is a given value of the source functions at the given point
X0.
One can remark that the time-average temperature measurement (6.5) represents a
non-local condition/measurement. It is convenient to use in practical situations where
a local measurement of the temperature at a fixed timeT1 ∈ (0, T ], namely,
u(x, T1) =: ψT1(x), x ∈ [0, L]
contains a large amount of noise. This may be due to harsh external conditions, or
to the fact that many space measurements can, in fact, never be recorded at a fixed
instant instantaneously. If this is the case, one can have a selection of such large noise
local temperature measurements, but which on average produce a less noisy nonlocal
measurement (6.5).
The individual separate cases concerning the identification of a single time-dependent
heat sourcer(t), whens(x) is known, or the identification of a single space-dependent
heat sources(x), whenr(t) is known, have been theoretically investigated in Prilepko
and Solov’ev [44] and Prilepko and Tkachenko [45], respectively.
The unique solvability of this inverse source problem was already established by
Ivanchov [26] and it is the objective of this study to obtain astable numerical solution
of this still ill-posed problem. Note that condition (6.6) was omitted in [26], but it
should be included in order to avoid non-uniqueness and instability cases which have
been given as counterexamples in [20]. For the inverse problem (6.1)–(6.6) we have
Chapter 6. 124
the following local unique solvability theorem.
Theorem 6.2.1 Assume that the following conditions are satisfied:
(i) u0(x), ψ(x) ∈ H2+γ[0, L], µ0(t), µL(t), χ(t) ∈ H1+γ/2[0, T ], h(x, t) ∈ Hγ,γ/2(DT )
with γ ∈ (0, 1),
f independent oft andf(x) ∈ Hγ[0, L],
g independent ofx andg(t) ∈ Hγ/2[0, T ];
(ii) f(X0) 6= 0,∫ T
0
g(t) dt 6= 0,g(t)
∫ T
0g(τ) dτ
≥ 0, ∀t ∈ [0, T ];
(iii) u0(0) = µ0(0), u0(L) = µL(0), u0(X0) = χ(0),∫ T
0
χ(t) dt = ψ(X0), ψ(0) =∫ T
0
µ0(t) dt, ψ(L) =∫ T
0
µL(t) dt.
Then for sufficiently smallT > 0 there exists a unique solution(r(t), s(x), u(x, t)) ∈Hγ/2[0, T ]×Hγ[0, L]×H2+γ,1+γ/2(DT ) of the inverse problem(6.1)–(6.6).
In this theorem the functions are required to lie in Holder spaces, defined as fol-
lows:
• H i+γ1,j+γ2(DT ), with i, j = 0, 1, 2 and0 < γ1, γ2 ≤ 1, denotes the space of
continuous functions withith partial derivative with respect tox andjth partial
derivative with respect tot such that there existm1 > 0 andm2 > 0 satisfying
|∂ixu(x1, t)−∂ixu(x2, t)| ≤ m1|x1−x2|γ1 , |∂jtu(x, t1)−∂jt u(x, t2)| ≤ m2|t1−t2|γ2
for all x1, x2 ∈ [0, L] andt1, t2 ∈ [0, T ].
• Hγ(Ω), with Ω = (0, L) or (0, T ), denotes the space of continuous functions
s : Ω → R with exponent0 < γ ≤ 1 such that there existsm > 0 satisfying
|s(x1)− s(x2)| ≤ m|x1 − x2|γ for all x1, x2 ∈ Ω.
In the next Sections 6.3 and 6.4, we will demonstrate how to solve the inverse heat
source problem (6.1)–(6.6) by using a regularised BEM.
Chapter 6. 125
6.3 The boundary element method (BEM)
In the numerical process, we utilise the BEM as introduced inSection 1.3 to the heat
conduction equation (6.1). We then obtain the following boundary integral equation
η(x)u(x, t)
=
∫ t
0
[
G(x, t, ξ, τ)∂u
∂n(ξ)(ξ, τ)− u(ξ, τ)
∂G
∂n(ξ)(x, t, ξ, τ)
]
ξ∈0,L
dτ
+
∫ L
0
G(x, t, y, 0)u(y, 0) dy+
∫ L
0
∫ T
0
G(x, t, y, τ)r(τ)f(y, τ) dτdy
+
∫ L
0
∫ T
0
G(x, t, y, τ)s(y)g(y, τ) dτdy+
∫ L
0
∫ T
0
G(x, t, y, τ)h(y, τ) dτdy,
(x, t) ∈ [0, L]× (0, T ]. (6.7)
Using the same discretisation as described in the previous chapters, we obtain
η(x)u(x, t) =N∑
j=1
[A0j(x, t)q0j + ALj(x, t)qLj − B0j(x, t)h0j − BLj(x, t)hLj ]
+
N0∑
k=1
Ck(x, t)u0,k + d1(x, t) + d2(x, t) + d0(x, t). (6.8)
where
d1(x, t) =
∫ L
0
∫ T
0
G(x, t, y, τ)r(τ)f(y, τ) dτdy, (6.9)
d2(x, t) =
∫ L
0
∫ T
0
G(x, t, y, τ)s(y)g(y, τ) dτdy, (6.10)
and can be calculated by applying the piecewise constant approximations to the func-
tionsf(x, t) andr(t) as the same in (2.21), and the functionsg(x, t) ands(x) as as
g(x, t) = g(xk, t), s(x) = s(xk) =: sk, (6.11)
Chapter 6. 126
for x ∈ (xk−1, xk], k = 1, N0. Then the double integrals (6.9) and (6.10) can be
approximated as
d1(x, t) =
∫ T
0
r(τ)
∫ L
0
G(x, t, y, τ)f(y, τ) dydτ =N∑
j=1
D1,j(x, t)rj ,
d2(x, t) =
∫ L
0
s(y)
∫ T
0
G(x, t, y, τ)g(y, τ) dτdy =
N0∑
k=1
D2,k(x, t)sk,
where
D1,j(x, t) =
∫ L
0
f(y, tj)Ayj(x, t) dy,
D2,k(x, t) =1
2
∫ T
0
g(xk, t)H(t− τ)
[
erf
(
x− xk−1
2√t− τ
)
− erf
(
x− xk
2√t− τ
)]
dτ,
These integrals are evaluated using Simpson’s rule for numerical integration. With
these approximations, the integral equation (6.8) becomes
η(x)u(x, t) =
N∑
j=1
[A0j(x, t)q0j + ALj(x, t)qLj − B0j(x, t)h0j − BLj(x, t)hLj]
+
N0∑
k=1
Ck(x, t)u0,k +N∑
j=1
D1,j(x, t)rj +
N0∑
k=1
D2,k(x, t)sk
+
N∑
j=1
D0,j(x, t). (6.12)
Applying the equation (6.12) at the boundary nodes(0, ti) and (L, ti) for i = 1, N
yields the system of2N linear equations
Aq¯−Bh
¯+ Cu
¯0+D1r
¯+D2s
¯+ d
¯= 0
¯, (6.13)
Chapter 6. 127
where matricesA, B, C and vectors q¯, h
¯, u
¯0, d
¯are defined the same as in (2.25), and
D1 =
D1,j(0, ti)
D1,j(L, ti)
2N×N
, D2 =
D2,k(0, ti)
D2,k(L, ti)
2N×N0
, s¯=[
sk
]
N0
.
Here, the boundary temperature h¯
is known by the boundary condition (6.3), i.e.
h¯=
u(0, tj)
u(L, tj)
2N
=
µ¯0µ¯L
2N
=: µ¯. (6.14)
Therefore from (6.13), we obtain
q¯= A−1
(
Bµ¯− Cu
¯0−D1r
¯−D2s
¯− d
¯
)
. (6.15)
To determine r¯
and s¯, the conditions (6.4)–(6.6) are imposed. Firstly, we consider
the interior points(X0, ti) for i = 1, N which can be written as
χi := χ(ti) = u(X0, ti), i = 1, N. (6.16)
Applying the interior points above to the equation (6.12) can give rise to the following
linear system ofN equations:
AIq¯−BIµ
¯+ CIu
¯0+DI
1r¯+DI
2s¯+ d
¯I = χ
¯, (6.17)
where
AI =[
A0j(X0, ti) ALj(X0, ti)]
N×2N, BI =
[
B0j(X0, ti) BLj(X0, ti)]
N×2N,
CI =[
Ck(X0, ti)]
N×N0
, DI1 =
[
D1,j(X0, ti)]
N×N, DI
2 =[
D2,k(X0, ti)]
N×N0
,
d¯I =
[
∑Nj=1D0,j(X0, ti)
]
N, χ
¯=[
χi
]
N.
Whereas the time-integral condition (6.5) is approximatedby the midpoint numerical
Chapter 6. 128
integration, as in the parallel way as (3.11), atx = xk for k = 1, N0, then we obtain
T
N
N∑
i=1
u(xk, ti) =
∫ T
0
u(xk, t) dt = ψ(xk) =: ψk for k = 1, N0. (6.18)
Using (6.12), equation (6.18) yields
T
N
N∑
i=1
[
AIIi q¯− BII
i µ¯+ CII
i u¯0
+DII1,ir¯
+DII2,is¯
+ d¯IIi
]
= ψ¯, (6.19)
where
AIIi =[
A0j(xk, ti) ALj(xk, ti)]
N0×2N, BII
i =[
B0j(xk, ti) BLj(xk, ti)]
N0×2N,
CIIi =
[
Ck(xk, ti)]
N0×N0
, DII1,i =
[
D1,j(xk, ti)]
N0×N, DII
2,i =[
D2,k(xk, ti)]
N0×N0
,
d¯IIi =
[
∑Nj=1D0,j(xk, ti)
]
N0
, ψ¯=[
ψk
]
N0
.
Finally, we consider the condition (6.6). Since we have usedthe space midpoint
discretisation, we then approximateS0 at the given pointX0 ∈ (0, L) as
S0 = s(X0) ≈s(xk∗) + s(xk∗+1)
2, (6.20)
where indexk∗ ∈ 1, . . . , N0 − 1 satisfiesxk∗ ≤ X0 < xk∗+1.
Now the approximate solutions r¯
and s¯
can be found by eliminating q¯
from (6.13)
and combining expressions (6.17), (6.19), and (6.20), to obtain, after some manipula-
tions, a linear system of(N +N0+1) equations with(N +N0) unknowns as follows:
Xw¯= y
¯, (6.21)
Chapter 6. 129
where
X =
AIA−1D1 −DI1 AIA−1D2 −DI
2
T
N
N∑
i=1
(AIIi A−1D1 −DII
1,i)T
N
N∑
i=1
(AIIi A−1D2 −DII
2,i)
0 . . . 0 0 . . . 0 12
120 . . . 0
(N+N0+1)×(N+N0)
,
y¯=
−χ+ AIA−1(Bµ¯− Cu
¯0− d
¯)− BIµ
¯+ CIu
¯0+ d
¯I
−ψ +T
N
N∑
i=1
(
AIIi A−1(Bµ
¯− Cu
¯0− d
¯)− BII
i µ¯+ CII
i u¯0
+ d¯IIi
)
S0
N+N0+1
,
and w¯=
r¯s¯
N+N0
.
Since the problem is ill-posed, then the system of equations(6.21) is ill-conditioned.
In the next section, we will deal with this ill-conditioningusing regularisation in order
to obtain a stable solution.
6.4 Regularisation
In practice, the measured data is unavoidably contaminatedby unplanned error. In
order to model this, we add noise into the input functionsχ(t) andψ(x) representing
the over-determination conditions (6.4) and (6.5) as follows:
χ¯ǫ = χ
¯+ random(′Normal′, 0, σχ, 1, N), (6.22)
and
ψ¯ǫ = ψ
¯+ random(′Normal′, 0, σψ, 1, N0), (6.23)
with the standard deviationsσχ andσψ to be taken as
σχ = p× maxt∈[0,T ]
|χ(t)|, and σψ = p× maxx∈[0,L]
|ψ(x)|, (6.24)
Chapter 6. 130
respectively. Note that the measurement (6.6) is already contaminated by error due to
the approximation made in (6.20).
If we consider the contamination of the right-hand side of equation (6.21) as‖y¯ǫ−
y¯‖ ≈ ǫ, then the direct least-squares solution w
¯= (XTX)−1XTy
¯ǫ will be unstable. To
overcome this instability, regularisation method needs tobe utilised. In this study, we
employ either the TSVD or the Tikhonov regularisation methods.
We first consider the use of the TSVD method as a regularisation procedure. To
use this method we use the[U,Σ,V]=svds(X,Nt) command in MATLAB, as we have
used previously in Chapter 3. In order to indicate the appropriate truncation levelNt,
theL-curve criterion, the GCV method, and the discrepancy principle are utilised.
Alternatively, the Tikhonov regularisation is another wayof obtaining a stable so-
lution of the ill-conditioned system of equations (6.21) which based on minimising the
regularised linear least-squares objective function
‖Xw¯− y
¯ǫ‖2 + λr‖R(1)r‖2 + λs‖R(2)s‖2 (6.25)
whereR(1), R(2) are (differential) regularisation matrices corresponding to a regular-
isation parameterλr, λs > 0, respectively. Solving (6.25) one obtains the regularised
solution
w¯λr ,λs
=(
XTX +RTR)−1
XTy¯ǫ. (6.26)
where the matrixR represents a block matrix of upper-left subblockλrR(1) and lower-
right subblockλsR(2). Initially, we takeλ := λr = λs and consider theL-curve
criterion, the GCV method, and the discrepancy principle aschoices for indicating
the single regularisation parameterλ. Note that both theL-curve and the GCV are
heuristic methods because they do not require the knowledgeof the level of noiseǫ.
More rigorously, one can use the discrepancy principle, [40], which selectsλ such that
‖Xw¯λ
− y¯ǫ‖ ≈ ǫ. (6.27)
Chapter 6. 131
If we allow for general multiple regularisation parametersλr andλs in (6.25) then, for
their selection one could employ theL-surface criterion, [3], which plots the residual
‖Xw¯λr ,λs
− y¯ǫ‖ versus‖R(1)r‖ and‖R(2)s‖ for various values ofλr andλs.
6.5 Numerical examples and discussion
In this section, we present two benchmark test examples in order to test the accuracy
of the approximate solutions. We are using the RMSE forr(t) as defined in (2.49)
whereas the RMSE fors(x) can be defined as
RMSE(s(x)) =
√
√
√
√
L
N0
N0∑
k=1
(sexact(xk)− snumerical(xk))2. (6.28)
6.5.1 Example 1
In the first example, we consider a smooth benchmark test withT = L = 1, X0 = 12
and the input data
u0(x) = u(x, 0) = x2, µ0(t) = u(0, t) = 0, µL(t) = u(1, t) = et,
χ(t) = u(12, t) =
et
4, ψ(x) =
∫ 1
0
u(x, t) dt = x2(e− 1),
S0 = s(12) = 1, f(x, t) = ex, g(x, t) = t+ 1,
h(x, t) = (x2 − 2)et − t2ex − (t + 1) sin(πx).
(6.29)
One can check that the conditions of Theorem 6.2.1 are satisfied hence the inverse
source problem (6.1)–(6.6) with the data (6.29) has a uniquesolution. It can easily be
verified through direct substitution that this solution is given by
u(x, t) = x2et, r(t) = t2, s(x) = sin(πx). (6.30)
As mentioned in Section 6.2, the inverse heat source problem(6.1)–(6.6) is ill-
Chapter 6. 132
posed since small errors in the measured data (6.4)–(6.6) cause large errors in the
solution. In order to quantify the degree of ill-conditioning we calculate the condition
number of the matrixX. The condition numbers forN = N0 ∈ 20, 40, 80 and
X0 ∈ 14, 12, 34 are shown in Table 6.1. In addition, the normalised singularvalues of
the matrixX are displayed in Figure 6.1, and the rapidly decreasing values indicate that
the system of equations (6.21) is ill-conditioned. Lookingat the columns of Table 6.1
it can be seen that the condition number only slightly decreases asX0 increase, hence
we do not expect the numerical results to be significantly influenced by the choice of
X0 within some interval[14, 34] away from the end pointsx = 0 andx = L = 1. Of
course, asX0 gets closer to the boundary pointx = 0 orx = L then the specification of
the interval temperature measurement (6.4) resembles a heat flux prescription, namely
ux(0, t) = limX0ց0
u(X0, t)− u(0, t)
X0, or ux(L, t) = lim
X0րL
u(X0, t)− u(L, t)
X0 − L.
However, this newly generated inverse problem in which Cauchy data are specified at
x = 0 or x = L is not addressed herein and it is deferred to a future work.
0 20 40 60 80 100 120 140 16010
−5
10−4
10−3
10−2
10−1
100
Nor
mal
ised
Sin
gula
r V
alue
s
2N
Figure 6.1: The normalised singular values of matrixX for N = N0 ∈ 20, 40, 80andX0 ∈ 1
4(− · −), 1
2(· · · ), 3
4(−−−), for Example 1.
In what follows, the numerical results are illustrated for afixed discretisationN =
N0 = 40 andX0 =12.
Chapter 6. 133
Table 6.1: The condition numbers of the matrixX in equation (6.21) for variousN =N0 ∈ 20, 40, 80 andX0 ∈ 1
4, 12, 34, for Example 1.
N = N0 20 40 80X0 = 1/4 1.94E+3 9.07E+3 5.04E+4X0 = 1/2 1.97E+3 7.51E+3 4.06E+4X0 = 3/4 1.93E+3 6.50E+3 3.33E+4
Exact Data
We consider first the case of exact data, i.e.p = 0 in (6.24). We directly solve
the linear system of equations (6.21) with the untruncated SVD method, and display
the numerical solutions forr(t), s(x), ux(0, t), andux(1, t) in Figures 6.2(a)–6.2(d),
respectively. From these figures, it can be seen that the solutions for r(t) ands(x)
are inaccurate, but the fluxesux(0, t) andux(1, t) are stable and accurate with small
RMSEs of 9.32E-3 and 4.58E-2, respectively, see Table 6.2. This is somewhat to be
expected since the inverse problem is ill-posed. Hence, regularisation is required to
overcome this instability.
For this, we utilise the TSVD and the Tikhonov regularisation of orders zero, one,
and two. The selection method of the regularisation parameters is first considered.
TheL-curves of the TSVD and the Tikhonov regularisations are presented in Figures
6.3(a) and 6.4(a), respectively. It can be seen that there isnoL-shape obtained for ei-
ther the TSVD, ZOTR, or FOTR, whereas the SOTR shows more clearly anL-corner
at λL=1E-1. Alternatively, the GCV method is utilised as anotherchoice for the reg-
ularisation parameter, as shown in Figure 6.3(b). The minimum of the GCV function
suggestsNt = 56 to be the truncation number for the TSVD, whilst for the Tikhonov
regularisation which is displayed in Figure 6.4(b), the minima indicate the parameters
λGCV =1.0E-7, 1.2E-7, and 4.5E-8 for ZOTR, FOTR and SOTR, respectively. Note
that for the exact data,ǫ ≈ 0 and the discrepancy principle cannot be employed. With
the GCV selection for the regularisation parameters determined from Figures 6.3(b)
and 6.4(b), the TSVD and the Tikhonov regularisation results are shown in Figure 6.5.
Compared to Figure 6.2, one can see that the instability of the numerical solutions is
Chapter 6. 134
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
1.2
1.4
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 1−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
1.2
x
s(x
)(b)
0 0.2 0.4 0.6 0.8 1−0.02
−0.015
−0.01
−0.005
0
0.005
t
ux(0
,t)
(c)
0 0.2 0.4 0.6 0.8 12
2.5
3
3.5
4
4.5
5
5.5
t
ux(1
,t)
(d)
Figure 6.2: The analytical (—–) and numerical results(− · −) of (a) r(t), (b) s(x), (c)ux(0, t), and (d)ux(1, t) obtained using the SVD for exact data, for Example 1.
not alleviated. We then employ another choice of the regularisation parameter based on
theL-curve method. This suggestsλL=1E-1 for the SOTR displayed in Figure 6.4(b).
Then with this choice forλ we obtain the stable and accurate numerical results shown
in Figure 6.6 and Table 6.2.
Noisy Data
Next, the case of noise contamination with percentagep = 1% is considered by adding
random noise into the input functionsχ(t) andψ(x) in (6.29), as in (6.22) and (6.23),
respectively. It is of crucial importance to utilise the regularisation in this case, and se-
lecting the regularisation parameter is the first step of theregularisation process. Here,
Chapter 6. 135
10−4
10−3
10−2
10−1
100
4.5
4.7
4.9
5.1
5.3
Nt=2
Nt=3
Nt=4
Nt=5
Nt=6
Nt=7
Nt=8,..,37
Nt=38,..,46
Nt=47,..,80
‖Xw¯Nt
− y¯
ǫ‖
‖R
w ¯N
t‖
(a)
10 20 30 40 50 60 70 80
10−10
10−9
10−8
10−7
Nt = 56
Nt
GC
V(N
t)
Nt=20
Nt=80
(b)
Figure 6.3: (a) TheL-curve and (b) the GCV function obtained by the TSVD for exactdata, for Example 1.
10−4
10−3
10−2
10−1
100
101
10−5
10−4
10−3
10−2
10−1
100
101
λ = 1E−8
‖Xw¯λ
− y¯
ǫ‖
‖R
w ¯λ‖
λ = 1E−8
λ = 1E−1
(a)
10−9
10−8
10−7
10−6
10−5
10−11
10−10
10−9
λ = 4.5E−8
λ
GC
V(λ
)
λ = 1.2E−7
λ = 1.0E−7
(b)
Figure 6.4: (a) TheL-curve, and (b) the GCV function, obtained by the ZOTR(−·−),FOTR(· · · ), and SOTR(−−−) for exact data, withλ = λr = λs, for Example 1.
theL-curve method and the discrepancy principle are employed ascriteria for choos-
ing the regularisation parameters. These are displayed in Figures 6.7 and 6.8 using the
TSVD and the Tikhonov regularisation, respectively. The suggested parameters are
given in Table 6.2. Figure 6.9 presents all results obtainedusing the TSVD and the
Tikhonov regularisation of orders zero, one, and two with the regularisation parame-
ters suggested by the discrepancy principle, see Table 6.2.Looking more closely at
Figure 6.9(a), it can be seen that the approximate solutionsfor r(t) obtained by the
Chapter 6. 136
first- and the second-order Tikhonov regularisation are reasonably stable, whereas the
numerical solution fors(x), as shown in Figure 6.9(b) is rather in accurate.
We consider the second-order Tikhonov regularisation withthe regularisation pa-
rameter suggested by theL-curve methodλL=10 and obtain the results shown in Fig-
ure 6.11. After analyzing this numerical solution, it can beclearly observed that we
cannot obtain accurate solutions for bothr ands usingλr = λs. Therefore, the case
λr 6= λs is considered and theL-surfaces are shown in Figure 6.10. On the plane of
logarithm of residual norm,log ‖Xw¯λ
− y¯ǫ‖, versus logarithm of the second deriva-
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
1.2
1.4
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 1−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
1.2
x
s(x
)
(b)
0 0.2 0.4 0.6 0.8 1−0.012
−0.01
−0.008
−0.006
−0.004
−0.002
0
0.002
t
ux(0
,t)
(c)
0 0.2 0.4 0.6 0.8 12
2.5
3
3.5
4
4.5
5
5.5
t
ux(1
,t)
(d)
Figure 6.5: The analytical (—–) and numerical results of (a)r(t), (b)s(x), (c)ux(0, t),and (d)ux(1, t) obtained using the TSVD(− + −), ZOTR (− · −), FOTR (· · · ),and SOTR(−−−) with regularisation parameters suggested by the GCV function ofFigure 6.3(b) and 6.4(b) for exact data, for Example 1.
Chapter 6. 137
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
1.2
x
s(x
)
(b)
0 0.2 0.4 0.6 0.8 1−0.005
0
0.005
0.01
0.015
0.02
t
ux(0
,t)
(c)
0 0.2 0.4 0.6 0.8 12
2.5
3
3.5
4
4.5
5
5.5
t
ux(1
,t)
(d)
Figure 6.6: The analytical (—–) and numerical results(− · −) of (a) r(t), (b) s(x), (c)ux(0, t), and (d)ux(1, t) obtained using the SOTR with the regularisation parameterλL=1E-1 suggested by theL-curve of Figure 6.4(a) for exact data, for Example 1.
tive of r, log ‖R(1)rλr‖, forms anL-shaped corner atλr=1E+1, whileλs=1 is based
around the area of theL-corner on the plane oflog ‖Xw¯λ
− y¯ǫ‖ versuslog ‖R(2)rλs‖.
However, the numerical solution fors(x) obtained using the parametersλr,L = 10,
λs,L = 1 suggested by theL-surface method, is still inaccurate. We finally use the trial
and error process to seek out the appropriated regularisation parameters, and found that
regularisation parametersλr,opt=8 andλs,opt=5.2E-2 can yield an accurate and stable
numerical solution, see Figure 6.11. Nevertheless, more research has to be undertaken
in the future for the selection of appropriate multiple regularisation parameters, [12].
for completeness, the RMSE of all results which we have mentioned so far are detailed
Chapter 6. 138
10−2
10−1
100
100
101
102
103
N t=2N t
=3N t=5
N t=4
Nt=53,54
Nt=55,..,57
Nt=66,67
N t=41,..,43N t=44,..,48N t
=49
Nt=62,..,65
Nt=50,51
Nt=58,..,61
Nt=68,..,76
Nt=77,78N
t=79N
t=80
‖Xw¯Nt
− y¯
ǫ‖
‖R
w ¯N
t‖
Nt=6,..,40
(a)
0 10 20 30 40 50 60 70 8010
−2
10−1
100
101
ε = 0.11
Nt = 14
Nt
‖X
w ¯N
t−
y ¯ǫ‖
(b)
Figure 6.7: (a) TheL-curve and (b) the discrepancy principle obtained using theTSVDfor noisy inputp = 1%, for Example 1.
0.06 0.08 0.1 0.2 0.410
−6
10−4
10−2
100
102
104
λ = 1E+1λ = 1E+2
‖Xw¯λ
− y¯
ǫ‖
‖R
w ¯λ‖
λ = 1E−1λ = 1
λ = 1E−4λ = 1E−3
(a)
10−5
10−4
10−3
10−2
10−1
100
101
102
10−1
100
λ = 1.5
λ
‖X
w ¯λ−
y ¯ǫ‖
ε = 0.11
λ = 2.8E−2λ = 1.3E−3
(b)
Figure 6.8: (a) TheL-curve and (b) the discrepancy principle obtained using theZOTR(− · −), FOTR(· · · ), and SOTR(−−−) for noisy inputp = 1%, with λ = λr = λs,for Example 1.
in Table 6.2.
6.5.2 Example 2
In example 1, the case of smooth source functions has been investigated and it can
retrieved the instability with the use of BEM together with the regularisation based on
either the TSVD and the Tikhonov regularisation. In this example, we are considering
Chapter 6. 139
0 0.2 0.4 0.6 0.8 1−0.2
0
0.2
0.4
0.6
0.8
1
1.2
1.4
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
1.2
x
s(x
)
(b)
0 0.2 0.4 0.6 0.8 1−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
t
ux(0
,t)
(c)
0 0.2 0.4 0.6 0.8 11.5
2
2.5
3
3.5
4
4.5
5
5.5
6
t
ux(1
,t)
(d)
Figure 6.9: The analytical (—–) and numerical results of (a)r(t), (b)s(x), (c)ux(0, t),and (d)ux(1, t) obtained using the TSVD(− + −) with Nt = 14, and the ZOTR(− · −), FOTR(· · · ), SOTR(−−−) with regularisation parameters suggested by thediscrepancy principle of Figure 6.8(b) for noisy inputp = 1%, for Example 1.
Chapter 6. 140
−2.35 −2.3 −2.25 −2.2 −2.15 −2.1 −2.05 −2−20
−10
0
10
−20
−15
−10
−5
0
5
log(‖Xw¯λ
− y¯
ǫ‖)
log(‖R(2)s¯λ‖)
log(‖R
(1)r ¯λ‖)
(a)
−2.35 −2.3 −2.25 −2.2 −2.15 −2.1 −2.05 −2 −1.95−20
−15
−10
−5
0
5
log(‖Xw¯λ
− y¯
ǫ‖)
log(‖R
(1)r ¯λ‖)
λr=1E−5
λr=1E+1
(b)
−2.35 −2.3 −2.25 −2.2 −2.15 −2.1 −2.05 −2 −1.95−20
−15
−10
−5
0
5
log(‖Xw¯λ
− y¯
ǫ‖)
log(‖R
(2)s ¯λ‖)
λs=1E−5
λs=1
(c)
Figure 6.10: TheL-surface on (a) a three-dimensional plot, (b) plane oflog ‖Xw¯λ−y
¯ǫ‖
versuslog ‖R(1)r¯λ‖, and (c) plane oflog ‖Xw
¯λ− y
¯ǫ‖ versuslog ‖R(2)s
¯λ‖, obtained
using the SOTR for noisy inputp = 1%, for Example 1.
Chapter 6. 141
0 0.2 0.4 0.6 0.8 1−0.2
0
0.2
0.4
0.6
0.8
1
1.2
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
1.2
x
s(x
)
(b)
0 0.2 0.4 0.6 0.8 1−0.02
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
t
ux(0
,t)
(c)
0 0.2 0.4 0.6 0.8 11.5
2
2.5
3
3.5
4
4.5
5
5.5
t
ux(1
,t)
(d)
Figure 6.11: The analytical (—–) and numerical results of (a) r(t), (b) s(x), (c)ux(0, t), and (d)ux(1, t) obtained using the SOTR with regularisation parameters sug-gested by theL-curve criterionλL = λr = λs = 10 (− − −), theL-surface method(λr,L, λs,L)=(10,1)(−+−), and the trial and error(λr,opt, λs,opt)=(8,5.2E-2)(− ∗ −),for noisy inputp = 1%, for Example 1.
Chapter 6. 142
Table 6.2: The RMSE forr(t), s(x), ux(0, t), andux(1, t) obtained using the SVD,TSVD, ZOTR, FOTR, and SOTR, forp ∈ 0, 1%, for Example 1.
Method p ParameterRMSE
r(t) s(x) ux(0, t) ux(1, t)SVD 0 - 1.47E-1 2.55E-1 9.32E-3 4.58E-2
TSVD 0 Nt=56 1.17E-1 2.03E-1 2.94E-3 3.57E-2ZOTR 0 λGCV =1.0E-7 1.20E-1 2.02E-1 3.53E-3 3.65E-2FOTR 0 λGCV =1.2E-7 7.62E-2 1.70E-1 3.68E-3 4.35E-2SOTR 0 λGCV =4.5E-8 7.96E-2 1.85E-1 6.48E-3 4.70E-2SOTR 0 λL=1.0E-1 8.70E-3 2.81E-2 1.27E-2 3.09E-3
SVD 1% - 1.62E+1 1.01E+2 2.84 2.48E-1TSVD 1% Nt=14 2.04E-1 1.77E-1 2.29E-2 5.83E-2ZOTR 1% λdis=1.3E-3 1.87E-1 1.83E-1 2.32E-2 4.87E-2FOTR 1% λdis=2.8E-2 1.28E-1 3.50E-1 1.05E-1 7.91E-2SOTR 1% λdis=1.5 9.72E-2 2.65E-1 8.05E-2 5.58E-2SOTR 1% λL=10 1.61E-1 4.23E-1 1.20E-1 9.36E-2SOTR 1% λr=10,λs=1 7.93E-2 2.18E-1 6.81E-2 4.40E-2SOTR 1% λr=8,λs=5.2E-2 1.92E-3 5.34E-2 1.00E-2 6.39E-3
in more severe case with the non-smooth source functions. Let T = L = 1, X0 = 12
and the input data
u0(x) = µ0(t) = µL(t) = 0, S0 = s(12) =
1
4,
χ(t) = u(12, t) = t2 sin(1
4), ψ(x) =
∫ 1
0
u(x, t) dt =sin(x− x2)
3,
f(x, t) = x, g(x, t) = et,
h(x, t) = (2t+ t2(1− 2x)2) sin(x− x2) + 2t2 cos(x− x2)
−x|t− 12| − et|x− 3
4|.
(6.31)
Note that the input data (6.31) satisfy the conditions of Theorem 6.2.1 to ensure the
existence and uniqueness of solution of the inverse problem(6.1)–(6.5). In fact, the
exact solution is given by
u(x, t) = t2 sin(x− x2), r(t) = |t− 12|, s(x) = |x− 3
4|.
Chapter 6. 143
0 20 40 60 80 100 120 140 16010
−5
10−4
10−3
10−2
10−1
100
Cond = 1.54E+4N
orm
alis
ed S
ingu
lar
Val
ues
2N
Cond = 8.69E+4
Cond = 3.46E+3
Figure 6.12: The normalised singular values of matrixX for N = N0 = 20 (− · −),N = N0 = 40 (· · · ), andN = N0 = 80 (−−−), for Example 2.
This is a more severe test example than Example 1 since the source componentsr(t)
ands(x) are not smooth functions.
We have calculated the condition numbers of the matrixX and obtained the condi-
tion numbers 3.46E+3, 1.54E+4, and 8.69E+4 forN = N0 = 20, 40, and 80, respec-
tively. Moreover, the corresponding normalised singular values are shown in Figure
6.12. In Example 2, the condition numbers of the matrixX are not much different
from the condition numbers for Example 1. Then we expect to solve this inverse prob-
lem by using wither the TSVD or the Tikhonov regularisation as means to reduce the
instability of the solution. Here, we fixN = N0 = 40 andX0 =12.
Exact Data
First we have tried the TSVD, ZOTR, FOTR and SOTR with the regularisation pa-
rameter given by the GCV function. This yieldsNt = 65, λGCV =2.9E-8, 3.2E-8, and
8.3E-9, respectively. But we have found that the solutions for r(t) ands(x) are not
so accurate. We then considered theL-curve method for choosing the regularisation
parameter. Figures 6.13(a) and 6.13(b) display theL-curves for the TSVD and the
Tikhonov regularisation, respectively. The same as theL-curve in Example 1, anL-
shape is obtained only when using the SOTR with suggests anL-corner aroundλ=1E-4
to 1E-3. In particular, forλL=1E-4 we obtain the stable solutions presented in Figure
Chapter 6. 144
10−5
10−4
10−3
10−2
10−1
100
0.4
0.6
0.8
1.0
1.5
2.0
2.5
3.0
3.5
Nt=1
Nt=2,3
Nt=4
Nt=5,.,10
N t=11,..,24N t
=25,..,34N t=35,..,41
N t=42,..,45
N t=46,..,4
9
N t=51,..,5
4
N t=56,..,5
9
N t=60,..,8
0
‖Xw¯Nt
− y¯
ǫ‖
‖R
w ¯N
t‖
(a)
10−4
10−3
10−2
10−1
100
10−6
10−5
10−4
10−3
10−2
10−1
100
101
λ = 1E−5
λ = 1E−4
λ = 1E−3λ = 1E−2
λ = 1E−1
λ = 1
λ = 1E+1
λ = 1E+2
‖Xw¯λ
− y¯
ǫ‖
‖R
w ¯λ‖
λ = 1E−5 λ = 1E−4 λ = 1E−3 λ = 1E−2λ = 1E−1
λ = 1
λ = 1E+1
λ = 1E+2
λ = 1E−4λ = 1E−3λ = 1E−2λ = 1E−1
λ = 1
λ = 1E+1
λ = 1E+2
(b)
Figure 6.13: TheL-curve obtained using (a) the TSVD and (b) the ZOTR(− · −),FOTR(· · · ), and SOTR(−−−) with λ = λr = λs, for exact data, for Example 2.
6.14 and Table 6.3. The untruncated SVD, i.e.Nt = 80, whose numerical results are
also included is not so accurate and stable in retrieving thefunctionsr(t) ands(x).
Table 6.3: The RMSE forr(t), s(x), ux(0, t), andux(1, t) obtained using the SVD,TSVD, ZOTR, FOTR, and SOTR, forp ∈ 0, 1%, for Example 2.
Method p ParameterRMSE
r(t) s(x) ux(0, t) ux(1, t)SVD 0 - 1.15E-1 4.12E-2 2.05E-3 6.95E-4
TSVD 0 Nt=65 1.17E-1 7.92E-2 1.15E-1 4.12E-2ZOTR 0 λGCV =2.9E-8 1.67E-1 6.63E-2 5.21E-3 2.19E-3FOTR 0 λGCV =3.2E-8 8.94E-2 2.99E-2 2.12E-3 6.01E-4SOTR 0 λGCV =8.3E-9 9.20E-2 3.13E-2 2.10E-3 7.61E-4SOTR 0 λL=1.0E-4 5.88E-3 8.94E-3 2.39E-3 1.04E-3
SVD 1% - 5.31E+1 8.91E+1 2.88 2.42E-1TSVD 1% Nt=10 2.16E-1 2.37E-1 1.15E-1 9.05E-2ZOTR 1% λdis=7.3E-4 1.20E-1 2.12E-1 1.15E-1 4.72E-2FOTR 1% λdis=3.2E-2 1.66E-1 6.77E-2 2.21E-2 2.51E-2SOTR 1% λdis=2.3 5.78E-2 7.90E-2 9.98E-3 4.95E-2SOTR 1% λL=1 9.24E-2 6.63E-2 1.55E-2 4.04E-2SOTR 1% λr,L=1,λs,L=10 3.96E-2 1.13E-1 4.67E-3 6.40E-2SOTR 1% λr,opt=2.2,λs,opt=5.9 2.37E-2 1.01E-1 3.42E-3 5.98E-2
Chapter 6. 145
0 0.2 0.4 0.6 0.8 1−0.2
−0.1
0
0.1
0.2
0.3
0.4
0.5
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
x
s(x
)
(b)
0 0.2 0.4 0.6 0.8 1−0.2
0
0.2
0.4
0.6
0.8
1
t
ux(0
,t)
(c)
0 0.2 0.4 0.6 0.8 1−1
−0.9
−0.8
−0.7
−0.6
−0.5
−0.4
−0.3
−0.2
−0.1
0
t
ux(1
,t)
(d)
Figure 6.14: The analytical (—–) and numerical results of (a) r(t), (b) s(x), (c)ux(0, t), and (d)ux(1, t) obtained using the SVD(− · −) and the SOTR(−o−) withthe regularisation parameterλL=1E-4 suggested by theL-curve of Figure 6.13(b) forexact data, for Example 2.
Noisy Data
When noise is present in the measured dataχ(t) andψ(x), the regularisation with
an appropriate parameter has to be carefully considered. Here we have tried solving
the perturbed problem withp = 1% noisy input by using the TSVD, ZOTR, FOTR,
and SOTR, with the regularisation parameter given by the discrepancy principle. This
yieldsNt = 10, λdis=7.3E-4, 3.2E-2, and 2.3, respectively. Although the discrepancy
principle is a rigorous method which uses the knowledge of noise, the RMSE errors
displayed in Table 6.3 are quite large. Alternatively, we consider theL-curve method
Chapter 6. 146
10−2
10−1
100
10−1
100
101
102
Nt=1
Nt=2,..,4N
t=5,..,30
Nt=31,..,45
Nt=46,..,49
Nt=50,..,64
Nt=65,66
Nt=67,..,75
Nt=76,77
Nt=78,79N
t=80
‖Xw¯Nt
− y¯
ǫ‖
‖R
w ¯N
t‖
(a)
0.08 0.1 0.210
−6
10−4
10−2
100
102
104
λ = 1E−2
λ = 1
λ = 1E+1
‖Xw¯λ
− y¯
ǫ‖
‖R
w ¯λ‖
λ = 1E−3λ = 1E−4
λ = 1E−1
(b)
Figure 6.15: TheL-curve obtained using (a) the TSVD and (b) the ZOTR(− · −),FOTR (· · · ), and SOTR(− − −) with λ = λr = λs, for noisy inputp = 1%, forExample 2.
for the choice of regularisation parameter displayed in Figure 6.15. This suggests the
appropriate parameters asNt between 5 and 30,λL=1E-4, 1E-2, and 1 for the TSVD,
ZOTR, FOTR, and SOTR,, respectively. We then solved the inverse problem with these
parameters and found that the numerical results obtained using the TSVD, ZOTR and
FOTR, are not so accurate. Whereas the SOTR yields a more accurate solution, as
shown in Figure 6.17 with dashed line. Hence, as in Example 1,the case ofλr 6= λs
needs to be considered by using theL-surface method for choosing the appropriate
regularisation parameters. Figures 6.16 displays theL-surface which selectsλr,L=10
andλs,L=1, and the results obtained using the SOTR with these parameters are shown
in Figure 6.17. Furthermore, the regularisation parameters selected by the trial and
error have also been considered and these results have also been included in Figure
6.17. The accurate retrieval ofr(t) is possible, but fors(x) this is less accurate.
6.6 Conclusions
This chapter has presented a numerical approach to the simultaneous numerical de-
termination of the space- and the time-dependent coefficient source functions of an
Chapter 6. 147
−2.25−2.2
−2.15−2.1
−20−15
−10−5
0−20
−15
−10
−5
0
log(‖Xw¯λ
− y¯
ǫ‖)log(‖R(2)s
¯λ‖)
log(‖R
(1)r ¯λ‖)
(a)
−2.25 −2.2 −2.15 −2.1−20
−18
−16
−14
−12
−10
−8
−6
−4
−2
0
log(‖Xw¯λ
− y¯
ǫ‖)
log(‖R
(1)r ¯λ‖)
λr=1E−3
λr=1
λr=10
(b)
−2.25 −2.2 −2.15 −2.1−20
−18
−16
−14
−12
−10
−8
−6
−4
−2
0
log(‖Xw¯λ
− y¯
ǫ‖)
log(‖R
(2)s ¯λ‖)
λs=1E−3
λs=1
λs=10
(c)
Figure 6.16: TheL-surface on (a) a three-dimensional plot, (b) plane oflog ‖Xw¯λ
−y¯ǫ‖
versuslog ‖R(1)r¯λ‖, and (c) plane oflog ‖Xw
¯λ− y
¯ǫ‖ versuslog ‖R(2)s
¯λ‖, obtained
using the SOTR for noisy inputp = 1%, for Example 2.
inverse heat conduction problem with Dirichlet boundary conditions together with
specified interior temperature measurement and time-integral condition, as the over-
determination conditions.
The numerical discretisation was based on the BEM together with either the TSVD,
or the Tikhonov regularisation. Additionally, various methods for choosing the regu-
larisation parameters have been utilised. The numerical results presented show that
accurate and stable numerical solutions can be achieved provided that the regularisa-
tion parameters are appropriately selected. The two-parameter selection has proved
to be difficult, as some of our numerical results obtained using several criteria, e.g.
Chapter 6. 148
0 0.2 0.4 0.6 0.8 1−0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
t
r(t
)
(a)
0 0.2 0.4 0.6 0.8 1−0.2
0
0.2
0.4
0.6
0.8
1
x
s(x
)(b)
0 0.2 0.4 0.6 0.8 1−0.2
0
0.2
0.4
0.6
0.8
1
1.2
t
ux(0
,t)
(c)
0 0.2 0.4 0.6 0.8 1−1
−0.8
−0.6
−0.4
−0.2
0
0.2
t
ux(1
,t)
(d)
Figure 6.17: The analytical (—–) and numerical results of (a) r(t), (b) s(x), (c)ux(0, t), and (d)ux(1, t) obtained using the SOTR with regularisation parameters sug-gested by theL-curve criterionλL = λr = λs = 1 (− − −), theL-surface method(λr,L, λs,L)=(1,10)(− + −), and the trial and error(λr,opt, λs,opt)=(2.2,5.9)(− ∗ −),for noisy inputp = 1%, for Example 2.
discrepancy principle, GCV,L-curve,L-surface, have shown. Nevertheless, more re-
search has to be undertaken in the future for the selection ofmultiple regularisation
parameters, [12].
In the next chapter we will consider reconstructing multiplicative space- and time-
dependent heat sources.
Chapter 7
Determination of Multiplicative Space-
and Time-dependent Heat Sources
7.1 Introduction
In the previous chapter, we have investigated the reconstruction of an additive source
of the formr(t)f(x, t)+ s(x)g(x, t). In this chapter, we consider the reconstruction of
a multiplicative source of the formr(t)s(x), in which bothr(t) ands(x) are unknown
functions. In contrast to the previously investigated linear reconstruction of the additive
source, Chapter 6, this new inverse source problem formulation is more difficult to
solve because it now becomes nonlinear. Moreover, its ill-posedness with respect to
small errors in the input data being blown up in the output source solution adds even
further difficulty.
The existence and uniqueness of the sourcesr(t), s(x) and the temperatureu(x, t)
of the inverse problem were already established in [47]. In this chapter, we consider
obtaining a stable solution by using the BEM together with a nonlinear minimisation.
The plan of the chapter is as follows. In Section 7.2, we give the mathematical for-
mulation of the inverse multiplicative source problem and state its unique solvability.
In Section 7.3, we describe the numerical discretisation ofthe problem based on the
149
Chapter 7. 150
BEM, whilst in Section 7.4 we introduce the inverse method for obtaining the solution
based on a nonlinear least-squares minimisation. Section 7.5 presents and discusses
numerical results and illustrates the need for employing regularisation in order to sta-
bilise the solution. Finally, Section 7.6 presents the conclusions of the study.
7.2 Mathematical formulation
Consider the following inverse initial-boundary value problem of finding the temper-
atureu(x, t) and the multiplicatively separable source functionF (x, t) := r(t)s(x)
satisfying the heat equation
ut = uxx + r(t)s(x), (x, t) ∈ DT , (7.1)
subject to the initial condition (1.7), namely
u(x, 0) = u0(x), x ∈ [0, L], (7.2)
the homogeneous Neumann boundary conditions
ux(0, t) = ux(L, t) = 0, t ∈ [0, T ], (7.3)
the additional temperature measurement
u(X0, t) = χ(t), t ∈ [0, T ], (7.4)
at a fixed sensor locationX0 ∈ (0, L), and
u(x, T ) = β(x), x ∈ [0, L], (7.5)
at the ‘upper-base’ final timet = T . Conditions (7.3) express that the ends0, L of
the finite slab(0, L) are insulated. In order to avoid trivial non-uniqueness represented
Chapter 7. 151
by the identityr(t)s(x) = r(t)c
· cs(x), with c arbitrary non-zero constant, we impose a
fixing condition, say
s(X0) = S0. (7.6)
In the above setting, the functionsu0, χ, β and the constantS0 are given. We further
assume that the conditions (7.2)–(7.5) are consistent, i.e. the following compatibility
conditions are satisfied:
u′0(0) = u′0(L) = β ′(0) = β ′(L) = 0, χ(0) = u0(X0), χ(T ) = β(X0). (7.7)
The unique solvability, i.e. existence and uniqueness of the solution of the inverse
problem (7.1)–(7.6), was established in [47]. With some slight corrections, this theo-
rem reads as follows.
Theorem 7.2.1 Suppose thatu0(x), β(x) ∈ W42(0, L), andχ(t) ∈ W
22(0, T ) satisfy
(7.7)and thatS0 6= 0. Also, assume that:
(i) M := χ′(0)− u′′0(X0) 6= 0, m :=χ′(T )− β ′′(X0)
M6= 0,
(ii) u′′′0 (0) = u′′′0 (L) = β ′′′(0) = β ′′′(L) = 0,
(iii) λ1 < 1, 4λ2λ3 − (1− λ1)2 ≤ 0, λ4 < 1,
where
λ1 :=2
m2M2max
M2 +4L2‖χ′′‖2
π2, 4L‖θ‖2 + Lm2‖u′′′0 ‖2
,
λ2 :=2
M2max
4L6
π4m4, 1
, λ3 :=2‖θ′‖2m2
+4‖χ′′‖2M2
(
2‖θ‖2m2
+ ‖u′′′0 ‖2)
,
λ4 :=1
m2M2max M2 + 2L3z + 4L2‖χ′′‖2, 4L3z + 4L3‖θ′‖2 + 2Lm2‖u′′′0 ‖2,
θ(x) = β ′′′(x)−mu′′′0 (x), z =1− λ12λ2
.
Then the inverse problem given by equations(7.1)–(7.6)has a unique solutionr(t) ∈W
12(0, T ), s(x) ∈ W
22(0, L) and
u(x, t) ∈ W4,22 (DT ) ∩ C(0, T ;W4
2(0, L)) ∩ C(0, L;W22(0, T )).
Chapter 7. 152
Note that in the above theorem,Wk2(Ω), with k ∈ 1, 2, 4 andΩ = (0, L) or
(0, T ), denotes the Sobolev space of functions consisting of all elements ofL2(Ω)
having generalised derivatives up to orderk inclusively inL2(Ω). Also, we denote
W4,22 (DT ) := u ∈ L2(DT )|∂jxu ∈ L2(0, L), j = 1, 4, and∂itu ∈ L2(0, T ), i = 1, 2.
Finally, C(0, T ;W42(0, L)) denotes the space of continuous mappings from(0, T ) to
W42(0, L) andC(0, L;W2
2(0, T )) denotes the space of continuous mappings from(0, L)
to W22(0, T ). The norms‖χ′′‖ and‖u′′′0 ‖ are understood inL2(0, T ) andL2(0, L),
respectively. Also, the norms ofθ andθ′ are inL2(0, L).
Although the inverse problem (7.1)–(7.6) has a unique solution it is still ill-posed
because it violates the continuous dependence upon the input data (7.4) and (7.5). In
the next section we will demonstrate how the BEM discretising numerically the heat
equation (7.1) can be used together with the regularisationin order to obtain a stable
solution.
7.3 The boundary element method (BEM)
In this section, we use the numerical procedure for discretising the inverse problem
(7.1)–(7.6) by using the BEM which results in the following boundary integral equa-
tion:
η(x)u(x, t) =
∫ t
0
[
G(x, t, ξ, τ)∂u
∂n(ξ)(ξ, τ)− u(ξ, τ)
∂G
∂n(ξ)(x, t, ξ, τ)
]
ξ∈0,L
dτ
+
∫ L
0
G(x, t, y, 0)u(y, 0) dy+
∫ L
0
∫ T
0
G(x, t, y, τ)r(τ)s(y) dτdy,
(x, t) ∈ [0, L]× (0, T ]. (7.8)
Chapter 7. 153
Using the same discretisation as described in Section 2.3 which has been used so far
as in previous chapters, we obtain
η(x)u(x, t) =N∑
j=1
[A0j(x, t)q0j + ALj(x, t)qLj −B0j(x, t)h0j −BLj(x, t)hLj ]
+
N0∑
k=1
[Ck(x, t)u0,k +Drk(x, t)sk] , (7.9)
where
Drk(x, t) =
∫ xk
xk−1
∫ t
0
G(x, t, y, τ)r(τ) dτ dy =
N∑
j=1
dj,k(x, t)rj, (7.10)
wheredj,k(x, t) =∫ xkxk−1
∫ tjtj−1
G(x, t, y, τ) dτ dy for j = 1, N , k = 1, N0. The double
integral source termdj,k(x, t) can be evaluated analytically to be given by
dj,k(x, t) =
0 ; t ≤ tj−1,
J(x, t, xk−1, tj−1)− J(x, t, xk, tj−1)
+(x− xk−1)
2
4− (x− xk)
2
4; tj−1 < t ≤ tj , x ≤ xk−1,
J(x, t, xk−1, tj−1)− J(x, t, xk, tj−1)
−(x− xk−1)2
4− (x− xk)
2
4; tj−1 < t ≤ tj , xk−1 < x ≤ xk,
J(x, t, xk−1, tj−1)− J(x, t, xk, tj−1)
−(x− xk−1)2
4+
(x− xk)2
4; tj−1 < t ≤ tj , x > xk,
J(x, t, xk−1, tj−1)− J(x, t, xk, tj−1)
−J(x, t, xk−1, tj) + J(x, t, xk, tj) ; t > tj ,
Chapter 7. 154
where
J(x, t, xk, tj) =
(
(x− xk)2
4+t− tj2
)
erf
(
x− xk2√t− tj
)
+
√t− tj2√π
(x− xk) exp
(
−(x− xk)2
4(t− tj)
)
.
By applying (7.9) at the boundary element nodes(0, ti) and(L, ti) for i = 1, N and
the homogeneous Neumann boundary condition (7.3), i.e.q0j = qLj = 0, we obtain
the system of2N equations
−Bh¯+ Cu
¯0+Drs
¯= 0
¯, (7.11)
whereDr =
∑Nj=1 dj,k(0, ti)rj
∑Nj=1 dj,k(L, ti)rj
2N×N0
.
For the direct problem, we can find now the boundary temperaturesu(0, ti) and
u(L, ti) from (7.11) as
h¯= B−1(Cu
¯0+Drs
¯). (7.12)
Furthermore, the interior temperaturesu(X0, ti) for i = 1, N from the additional con-
dition condition (7.4) can be approximated similarly as in (6.16), i.e. [u(X0, ti)]N =[
χ(ti)]
N. Applying this in (7.9) it gives
−BIh¯+ CIu
¯0+DrIs
¯= χ
¯, (7.13)
whereDrI =[
∑Nj=1 dj,k(X0, ti)rj
]
N×N0
. Whereas the final temperatureu(xk, T ) for
k = 1, N0 from the overdetermination (7.5) can be approximated as[u(xk, T )]N0=
[
β(xk)]
N0
. Applying this in (7.9) it gives
−BIIIh¯+ CIIIu
¯0+DrIIIs
¯= β
¯, (7.14)
Chapter 7. 155
where
BIII =[
B0j(xk, T ) BLj(xk, T )]
N0×2N, CIII =
[
Ck(xk, T )]
N0×N0
,
DrIII =[
∑Nj=1 dj,k(xk, T )rj
]
N0×N0
.
7.4 Solution of inverse problem
In this section, we wish to obtain simultaneously the unknown componentsr(t) and
s(x) of the multiplicative source term in the inverse problem (7.1)–(7.6) by using the
BEM together with a classical minimisation process. The conditions (7.4)–(7.6) are
imposed by minimising the nonlinear least-squares function
F0(r, s) :=
N∑
i=1
(
u(X0, ti)− χ(ti))2
+
N0∑
k=1
(u(xk, T )− β(xk))2 + (s(X0)− S0)
2.
(7.15)
Here, the approximated temperaturesu(X0, ti) andu(xk, T ), as introduced earlier in
(7.13) and (7.14), respectively, are now employed into the above objective function
with the initial guesses r¯0
and s¯0
for functionsr ands, respectively. Whereass(X0) is
approximated the same as in (6.20). Then, applying the approximations (7.12)–(7.14)
we obtain
F0(r¯, s¯) =‖ − BIB−1(Cu
¯0+Drs
¯) + CIu
¯0+DrIs
¯− χ
¯‖2
+ ‖ − BIIIB−1(Cu¯0
+Drs¯) + CIIIu
¯0+DrIIIs
¯− β
¯‖2
+ (s(X0)− S0)2, (7.16)
where r¯= (rj)N and s
¯= (sk)N0
. The minimisation of (7.16) is performed using the
lsqnonlinroutine from the MATLAB Optimisation Toolbox. This routineattempts to
find the minimum of a sum of squares by starting from some arbitrary initial guesses r¯0
and s¯0
. Note that we have compiled this routine with the following default parameters:
Chapter 7. 156
• Algorithm = Trust-Region-Reflective.
• Maximum number of objective function evaluations, ‘MaxFunEvals’ = 100 ×(N +N0 + 1).
• Maximum number of iterations, ‘MaxIter’ =400.
• Termination tolerance on the function value, ‘TolFun’ =10−10 to 10−6.
• Termination tolerance, ‘TolX’ =10−10 to 10−6.
Of course, finding a global minimiser to a nonlinear optimisation problem is not an
easy task since the functional (7.15), which is in general not convex, i.e. the Hessian
of F is not positive definite. As a consequence it may have local minima in which a
descent method tends to get stuck, if the underlying inverseproblem is ill-posed, [14,
p.17]. In the next section we shall elaborate more on the choice of the initial guess for
the iterative routine, as well as on incorporating regularisation in the functional (7.15)
in order to ensure convergence to the desired stable solution.
7.5 Numerical examples and discussion
This section presents three benchmark test examples in order to test the accuracy and
stability of the numerical methods introduced in Sections 7.3 and 7.4. The RMSEs
for r(t) ands(x), defined in (2.49) and (6.28), respectively, are used to evaluate the
accuracy of the numerical results.
7.5.1 Example 1
We consider a benchmark test example withT = 1, L = 1/10, X0 = 1/20, and the
initial data (7.2) given by
u0(x) = u(x, 0) = 0, x ∈ [0, L]. (7.17)
Chapter 7. 157
For the direct problem (7.1)–(7.3) we also need the input source data
r(t) = − et
40
(
400π2t2 − 400π2t + t2 + t− 1)
, s(x) = 40 cos(20πx). (7.18)
In order to test the BEM accuracy for the direct problem givenby equation (7.1) with
the source given by the product of the functions in (7.18), subject to the homogenous
Neumann boundary condition (7.3) and the initial condition(7.17), the numerical re-
sults are compared with the analytical solution given by
u(x, t) = et(t− t2) cos(20πx). (7.19)
The exact expressions for the inputs (7.4)–(7.6) are given by
χ(t) = u(1/20, t) = −(t− t2)et, β(x) = u(x, 1) = 0,
S0 = s(1/20) = −40.
(7.20)
As defined in Theorem 7.2.1, we then haveS0 = −40 6= 0,M = −1 6= 0,m = −e 6=0, θ(x) = β(x) = u0(x) ≡ 0, λ1 = 0.2962 < 1, λ2 = 2, λ3 = 0, 4λ2λ3 − (1− λ1)
2 =
−0.4953 ≤ 0, z = 0.1759, andλ4 = 0.2613 < 1 which satisfy all the conditions
(i)–(iii) for existence and uniqueness of the solution.
As the specified boundary conditions (7.3) are of Neumann type, the boundary
unknowns in the BEM are represented by the Dirichlet datau(0, t) andu(L, t), as given
by (7.12). Once all the boundary data has been obtained accurately, equations (7.13)
and (7.14) can be employed explicitly and with no need of interpolations to obtain
the temperaturesu( 120, ti) andu(xk, 1) for i = 1, N andk = 1, N0, respectively. The
RMSE of the direct problem results are shown in Table 7.1 and it can be concluded that
the BEM numerical solutions are convergent to their corresponding exact values, as the
number of boundary elements increases. Whereas Figure 7.1 displays the analytical
and numerical results ofχ(t) andβ(x) and very good agreement can be observed.
Next we consider the inverse problem given by equations (7.1), (7.3), (7.17) and
Chapter 7. 158
Table 7.1: The RMSE foru(0, t), u(0.1, t), χ(t) andβ(x), obtained using the BEM forthe direct problem withN = N0 ∈ 10, 20, 40, for Example 1.
N = N0RMSE
u(0, t) u(0.1, t) χ(t) β(x)10 5.01E-3 5.01E-3 5.64E-3 8.51E-220 1.03E-3 1.03E-3 1.75E-3 4.51E-240 8.17E-4 8.17E-4 9.69E-4 2.30E-2
0 0.2 0.4 0.6 0.8 1−0.45
−0.4
−0.35
−0.3
−0.25
−0.2
−0.15
−0.1
−0.05
0
t
χ(t
)
(a)
0 0.02 0.04 0.06 0.08 0.1−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
x
β(x
)
(b)
Figure 7.1: The analytical (—–) and numerical results for (a) χ(t) and (b) β(x)obtained using the BEM for the direct problem withN = N0 ∈ 10 (− ·−), 20 (· · · ), 40 (−−−), for Example 1.
(7.20). The numerical solution can be obtained, as described in Section 7.4, by min-
imising the objective function (7.15). Preliminary numerical investigations showed
that the initial guesses r¯0
and s¯0
cannot be so arbitrary in order for the minimisation
process to converge globally. After many trials, we decidedto illustrate the numerical
results obtained by considering the initial guess as
r¯0
= r¯+ random(′Normal′, 0, σr, N, 1),
s¯0
= s¯+ random(′Normal′, 0, σs, N0, 1),
(7.21)
with the standard deviationsσr andσs, respectively, given by
σr = p0 × maxt∈[0,T ]
|r(t)|, σs = p0 × maxx∈[0,L]
|s(x)|, (7.22)
Chapter 7. 159
wherep0 is a percentage of perturbation. Hereafter, unless otherwise specified, we
present results obtained withp0 = 100% perturbed initial guess (which is quite far
from the exact solution (7.18)) andN = N0 = 20, we also have set parameters TolFun
= TolX = 10−6 for the MATLAB optimisation toolboxlsqnonlinto solve the inverse
problem.
Figure 7.2(a) shows the unregularised objective functionF0 which converges in
39 iterations and the numerical results forr(t), s(x), u(0, t), u(0.1, t) are displayed
in Figures 7.2(b)–7.2(e), respectively. As we can see in these figures, the numerical
results are inaccurate and partially unstable in Figure 7.2(c).
In order to improve the accuracy and stability, we apply a Tikhonov regularisation
process based on minimising the penalised objective function
Fλ(r¯, s¯) := F0(r
¯, s¯) + λ
(
‖Rr‖2 + ‖Rs‖2)
, (7.23)
whereλ > 0 is a regularisation parameter to be prescribed, andR is a (differential)
regularising matrix as introduced in Section 1.6. Initially, we have applied the first-
and second-order regularisations based on minimising the objective function (7.23) as
Fλ(r¯, s¯) =F0(r
¯, s¯) + λ
(
N−1∑
i=1
(ri+1 − ri)2 +
N0−1∑
k=1
(sk+1 − sk)2
)
, (7.24)
Fλ(r¯, s¯) =F0(r
¯, s¯) + λ
(
N−1∑
i=2
(−ri+1 + 2ri − ri−1)2 +
N0−1∑
k=2
(−sk+1 + 2sk − sk−1)2
)
,
(7.25)
respectively.
By trial and error, among various regularisation parameters λ ∈ 10−9, . . . , 102,
we have found, as illustrative stable results, those obtained withλopt = 10−5 which
are shown in Figure 7.3. As we can see in this figure, applying orders one or two
regularisations (7.24) or (7.25) yield stable, but rather inaccurate results, especially
near the endpoints of the intervals of definition of the functions involved, see Figure
Chapter 7. 160
0 10 20 30 4010
−4
10−2
100
102
104
106
108
Number of iterations
Obje
ctiv
eFunct
ion
F0
39
(a)
0 0.2 0.4 0.6 0.8 1−100
−50
0
50
100
150
t
r(t
)
(b)
0 0.02 0.04 0.06 0.08 0.1−150
−100
−50
0
50
100
150
x
s(x
)
(c)
0 0.2 0.4 0.6 0.8 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
t
u(0
,t)
(d)
0 0.2 0.4 0.6 0.8 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
t
u(0
.1,t)
(e)
Figure 7.2: (a) The objective functionF0 and the numerical results for (b)r(t), (c)s(x), (d) u(0, t), (e) u(0.1, t) obtained with no regularisation(− · −), for exact datafor Example 1. The corresponding analytical solutions are shown by continuous line(—–) in (b)–(e) and thep0 = 100% perturbed initial guesses are shown by dotted line(· · · ) in (b) and (c).
Chapter 7. 161
0 0.2 0.4 0.6 0.8 10
10
20
30
40
50
t
r(t
)
(a)
0 0.02 0.04 0.06 0.08 0.1−60
−40
−20
0
20
40
60
x
s(x
)
(b)
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
t
u(0
,t)
(c)
0 0.2 0.4 0.6 0.8 1−0.05
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
t
u(0
.1,t
)
(d)
Figure 7.3: The numerical results for (a)r(t), (b)s(x), (c)u(0, t), (d)u(0.1, t) obtainedwith the first-order regularisation(· · · ) and the second-order regularisation(− − −)with regularisation parameterλopt = 10−5, for exact data for Example 1. The corre-sponding analytical solutions are shown by continuous line(—–).
7.3(b). In order to improve on these inaccuracies we have then investigated a hybrid
combination of first- and second- order regularisations given by
Fλ(r¯, s¯) = F0(r
¯, s¯) + λ
(
(r1 − r2)2 + (−rN−1 + rN)
2 +
N−1∑
i=2
(−ri+1 + 2ri − ri−1)2
+ (s1 − s2)2 + (−sN0−1 + sN0
)2 +
N0−1∑
k=2
(−sk+1 + 2sk − sk−1)2
)
. (7.26)
Chapter 7. 162
According to (7.23) and (7.26), the differential regularisation matrixR is given by
R =
1 −1 0 0 .
−1 2 −1 0 .
0 −1 2 −1 .
. . . . .
. (7.27)
In the regularisation process, we need to choose an appropriate regularisation parame-
terλ which balances accuracy and stability. Here, we use theL-curve method to find
the regularisation parameterλ. Figure 7.4(a) shows theL-curve obtained by plotting
the solution norm√
‖Rr‖2 + ‖Rs‖2 versus the residual norm√
F0(r¯, s¯) for various
values ofλ whenR is given by (7.27). From this figure it can be seen that the corner of
theL-curve occurs nearbyλL = 10−5, with other appropriate values between the wide
range10−6 to 10−4. With this value of the regularisation parameter, the regularised ob-
jective functionFλ and the numerical results are shown in Figures 7.4(b)–7.4(f). From
Figure 7.4(b) it can be seen that convergence for the regularised objective functionFλ
is achieved within 15 iterations. Also, in comparison with the previous Figures 7.3(a)–
7.3(d), very good agreement between the exact and the regularised numerical solutions
is now obtained, as illustrated in Figures 7.4(c)–7.4(f), respectively. All results are
summarised in terms of the RMSE in Table 7.2. Various initialguesses (7.21) with
p0 ∈ 40, 60, 80, 100% in (7.22) have been investigated in order to test the robustness
of the minimisation procedure with respect to the independence on the initial guess.
From Table 7.2 it can be seen that whilst the choice of the initial guess seems to matter
for the accuracy of the unregularised solution; i.e.λ = 0, this restriction disappears
when regularisation withλL = 10−5 is imposed. This shows that the numerical reg-
ularisation method employed is robust with respect to the independence on the initial
guess.
To test the stability of the BEM combined with the nonlinear regularisation, we
solve the inverse problem when random noises are added to theinput functionsχ(t)
Chapter 7. 163
10−2
10−1
100
101
0
20
40
60
80
100
120λ=10−9
λ=10−8
λ=10−7
λ=10−6
λ=10−5
λ=10−4λ=10−3
λ=10−2λ=10−1 λ=1
√
F0(r¯, s¯)
√
‖R
r ¯‖2
+‖R
s ¯‖2
(a)
0 5 10 1510
−4
10−2
100
102
104
106
108
Number of iterations
Obje
ctiv
eFunct
ion
Fλ
15
(b)
0 0.2 0.4 0.6 0.8 10
5
10
15
20
25
30
35
40
45
t
r(t
)
(c)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09−50
−40
−30
−20
−10
0
10
20
30
40
50
x
s(x
)
(d)
0 0.2 0.4 0.6 0.8 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
t
u(0
,t)
(e)
0 0.2 0.4 0.6 0.8 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
t
u(0
.1,t)
(f)
Figure 7.4: (a) TheL-curve criterion, (b) the objective functionFλ, and the numericalresults(− −) for (c) r(t), (d) s(x), (e)u(0, t), (f) u(0.1, t) obtained with the hybrid-order regularisation (7.26) with regularisation parameter λL = 10−5 suggested byL-curve, for exact data Example 1. The corresponding analytical solutions are shown bycontinuous line (—–) in (c)–(f).
Chapter 7. 164
0 2 4 6 8 10 12 14 1610
−4
10−2
100
102
104
106
108
Number of iterations
Obje
ctiv
eFunct
ion
Fλ
16
15
15
(a)
0 0.2 0.4 0.6 0.8 10
5
10
15
20
25
30
35
40
45
t
r(t
)
(b)
0 0.02 0.04 0.06 0.08 0.1−50
−40
−30
−20
−10
0
10
20
30
40
50
x
s(x
)
(c)
0 0.2 0.4 0.6 0.8 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
t
u(0
,t)
(d)
0 0.2 0.4 0.6 0.8 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
t
u(0
.1,t
)
(e)
Figure 7.5: (a) The objective functionFλ and the numerical results for (b)r(t), (c)s(x),(d) u(0, t), (e)u(0.1, t) obtained with the hybrid-order regularisation (7.26) withregu-larisation parameterλL = 10−5 suggested byL-curve forp ∈ 1(−·−), 3(· · · ), 5(−−−)% noisy data, for Example 1. The corresponding analytical solutions are shown bycontinuous line (—–) in (b)–(e).
Chapter 7. 165
Table 7.2: The RMSE forr(t), s(x), u(0, t), u(0.1, t) for exact data, Example 1.
p0(%)parameter RMSE
λ r(t) s(x) u(0, t) u(0.1, t)
40%0 6.349 18.47 4.72E-2 3.27E-2
λL=1E-5 1.528 0.819 1.49E-2 1.53E-2
60%0 9.752 26.70 1.14E-1 8.12E-2
λL=1E-5 1.513 0.767 1.48E-2 1.50E-2
80%0 25.83 44.93 1.99E-1 2.74E-1
λL=1E-5 1.526 0.812 1.48E-2 1.47E-2
100%0 53.70 54.24 2.34E-1 2.54E-1
λL=1E-5 1.529 0.819 1.47E-2 1.48E-2
andβ(x) as
χ¯ǫ = χ
¯+ random(′Normal′, 0, σχ, N, 1),
β¯ǫ = β
¯+ random(′Normal′, 0, σβ, N0, 1),
(7.28)
with the standard deviationsσχ andσβ given by
σχ = p× maxt∈[0,T ]
|χ(t)|, σβ = p× maxx∈[0,L]
|β(x)|, (7.29)
respectively. The numerical results obtained withλL = 10−5, are illustrated in Fig-
ure 7.5. From Figure 7.5(a) it can be seen that convergence ofthe hybrid-order reg-
ularised objective functional (7.26) is rapidly achieved within 15-16 iterations for
p ∈ 1, 3, 5%. Furthermore, Figures 7.5(b)–7.5(e) show that stable and accurate
numerical results are obtained for all amounts of noisep. Also, as expected, the nu-
merical solutions become more accurate as the amount of noisep decreases.
7.5.2 Example 2
In Example 1, all conditions for the existence and uniqueness of Theorem 7.2.1 were
satisfied. We now consider an example which has the analytical solution, [48],
u(x, t) = (e3t − e−t) cos(x), r(t) = e3t, s(x) = 4 cos(x), (7.30)
Chapter 7. 166
whereT = 0.3, L = π. One can easily check that the homogeneous Neumann con-
ditions (7.3) are satisfied and that the initial condition (7.2) is also homogeneous, as
given in (7.17). TakingX0 = 0.75 we obtain that the input data (7.4)–(7.6) are given
by
χ(t) = u(0.75, t) = (e3t − e−t) cos(0.75),
β(x) = u(x, 0.3) = (e0.9 − e−0.3) cos(x),
S0 = s(0.75) = 4 cos(0.75).
(7.31)
From this we haveS0 = M = 4 cos(0.75) 6= 0, u0(x) ≡ 0, m = e0.9 6= 0, θ(x) =
(e0.9 − e−0.3) sin(x), λ1 = 5.3719, λ2 = 0.2518, λ3 = 24.928, z = −8.6813, λ4 =
14.654. One can then observe that the conditions (i) and (ii) of Theorem 7.2.1 are
satisfied, but the condition (iii) has been violated. Whilsta solution obviously exists,
as given by equations (7.30), one cannot guarantee yet that this solution is unique.
We have solved first the direct problem given by equations (7.1) (with r ands given
by (7.30)), (7.3) and (7.17) using the BEM with various numbers of boundary elements
N = N0 ∈ 5, 10, 20 and the numerical results forχ(t) andβ(x) presented in Fig-
ure 7.6 show rapid convergence and excellent agreement withthe analytical solution
(7.31). Afterwards, we have solved the inverse problem given by equations (7.1), (7.3),
(7.17) and (7.31) in order to retrieve the temperatureu(x, t) and the heat source com-
ponentsr(t) ands(x) given analytically by (7.30). We have taken boundary elements
N = N0 = 20 and the arbitrary initial guesses r¯0
= 0¯
and s¯0
= 0¯.
We first consider the case of exact data. The convergence of the unregularised
objective functionF0 achieved within 56 iterations using thelsqnonlin routine with
TolFun = TolX =10−10 is illustrated in Figure 7.7(a). Also, the RMSEs of solutions r
ands are shown in Figure 7.7(b) by dash line(· · · ) and dot line(−−−), respectively.
The numerical solutions forr ands obtained after 56 iterations are shown by dash-dot
line (− · −) in Figures 7.7(c) and 7.7(d), respectively. Very good agreement between
the numerical and analytical solutions fors can be observed, whilst the numerical
solution forr is stable but slightly away from the analytical solution. Wethen look
Chapter 7. 167
0 0.05 0.1 0.15 0.2 0.25 0.3 0.350
0.2
0.4
0.6
0.8
1
1.2
1.4
t
χ(t
)
(a)
0 0.5 1 1.5 2 2.5 3 3.5−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
x
β(x
)
(b)
Figure 7.6: The analytical (—–) and numerical results for (a) χ(t) and (b)β(x)obtained using the BEM for the direct problem withN = N0 ∈ 5(− ·−), 10(· · · ), 20(−−−), for Example 2.
more closely at Figure 7.7(b) and observe that the minimum ofRMSEs is at iteration
31 instead of 56. Therefore, we have tried solving the inverse problem with the fixed
iteration at 31, and the numerical results become more accurate, as illustrated by the
circle markers() in Figures 7.7(c) and 7.7(d). Further, we have applied the hybrid-
order regularisation procedure (7.26) with the regularisation parameterλopt = 2 ×10−4 (chosen by the trial and error) and the results are shown in Figure 7.8. Figure
7.8(a) displays the convergence of the regularised functional (7.26) achieved within
28 iterations. Also, results for RMSEs and the solutions forr and s are shown in
Table 7.3: The RMSE forr(t) ands(x), for the noise levelsǫ0 ∈ 0, 0.01, 0.1, forExample 2.
Noise level No. of iterationsparameter RMSE
λ r(t) s(x)
No noise
56 0 6.3068E-2 1.5492E-131 (fixed) 0 2.6668E-2 1.4751E-1
28 λopt=2E-4 5.6471E-2 1.1003E-123 (fixed) λopt=2E-4 1.9713E-2 6.4412E-2
ǫ0 = 0.0127 λopt=4E-4 6.3004E-2 1.0886E-1
21 (fixed) λopt=4E-4 2.8829E-2 6.5281E-2ǫ0 = 0.1 17 λopt=2 5.2212E-2 3.0714E-2
Chapter 7. 168
Figures 7.8(b)–7.8(d). From Figure 7.8(b) one can see that the minimum of the RMSEs
occurs after 23 iterations. By comparing Figures 7.7 and 7.8one can conclude that the
inclusion of some small regularisation yields slightly more accurate and stable results.
Next, we consider the stability of the numerical solution when the noise is present
0 10 20 30 40 50 6010
−10
10−8
10−6
10−4
10−2
100
102
Number of iterations
Obje
ctiv
eFunct
ion
F0
56
31
(a)
0 10 20 30 40 50 6010
−2
10−1
100
101
Number of iterations
Root
Mea
nSquare
Err
or
(b)
0 0.05 0.1 0.15 0.2 0.25 0.30.8
1
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
t
r(t
)
(c)
0 0.5 1 1.5 2 2.5 3 3.5−5
−4
−3
−2
−1
0
1
2
3
4
5
x
s(x
)
(d)
Figure 7.7: (a) The objective functionF0, (b) the RMSEs forr(t) (− − −) ands(x)(· · · ) obtained with no regularisation for exact data, and the numerical results for (c)r(t) and (d)s(x) obtained using the minimisation process after 56 unfixed iterations(− · −), and 31 fixed iterations( ), for Example 2. The corresponding analyticalsolutions (7.30) are shown by continuous line (—–) in (c) and(d).
Chapter 7. 169
0 5 10 15 20 25 3010
−5
10−4
10−3
10−2
10−1
100
101
102
Number of iterations
Obje
ctiv
eFunct
ion
Fλ
2823
(a)
0 5 10 15 20 25 3010
−2
10−1
100
101
Number of iterations
Root
Mea
nSquare
Err
or
(b)
0 0.05 0.1 0.15 0.2 0.25 0.30.8
1
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
t
r(t
)
(c)
0 0.5 1 1.5 2 2.5 3 3.5−5
−4
−3
−2
−1
0
1
2
3
4
5
x
s(x
)
(d)
Figure 7.8: (a) The objective functionFλ, (b) the RMSEs forr(t) (− − −) ands(x)(· · · ) obtained using the hybrid-order regularisation (7.26) with regularisation param-eterλopt = 2 × 10−4 for exact data, and the numerical results for (c)r(t) and (d)s(x)obtained using minimisation process after 28 unfixed iterations(− · −), and 23 fixediterations( ), for Example 2. The corresponding analytical solutions (7.30) areshown by continuous line (—–) in (c) and (d).
in the input data (7.4) and (7.5). As in [48], the noise was defined by
χǫ(ti) = χ(ti)
1 +ǫ0
√
∑Ni=1 χ
2(ti)rand(i)
, i = 1, N,
βǫ(xk) = β(xk)
1 +ǫ0
√
∑Nk=1 β
2(xk)rand(k)
, k = 1, N0,
(7.32)
whererand(·) is a random variable generated by the MATLAB command from a nor-
Chapter 7. 170
0 5 10 15 20 25 3010
−4
10−3
10−2
10−1
100
101
102
Number of iterations
Obje
ctiv
eFunct
ion
Fλ
2127
(a)
0 5 10 15 20 25 3010
−2
10−1
100
101
Number of iterations
Root
Mea
nSquare
Err
or
(b)
0 0.05 0.1 0.15 0.2 0.25 0.30.8
1
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
t
r(t
)
(c)
0 0.5 1 1.5 2 2.5 3 3.5−5
−4
−3
−2
−1
0
1
2
3
4
5
x
s(x
)
(d)
Figure 7.9: (a) The objective functionFλ, (b) the RMSEs forr(t) (− − −) ands(x)(· · · ) obtained using the hybrid-order regularisation (7.26) with regularisation parame-terλopt = 4× 10−4 for noise levelǫ0 = 0.01, and the numerical results for (c)r(t) and(d) s(x) obtained using the minimisation process after 27 unfixed iterations(− · −),and 21 fixed iterations( ), for Example 2. The corresponding analytical solutions(7.30) are shown by continuous line (—–) in (c) and (d).
mal distribution with mean zero and unit standard deviation, and ǫ0 represents the
noise level. Remark that the noise (7.32) is multiplicative, whilst the noise in (7.28),
Example 1, is additive. Forǫ0 = 0.01, Figure 7.9 illustrates the results obtained using
the hybrid-order regularisation (7.26) with regularisation parameterλopt = 4 × 10−4.
The convergence of the regularised objective function achieved within 27 iterations is
shown in Figure 7.9(a), whilst the minimum RMSEs ofr ands occur after 21 itera-
tions, as can be seen in Figure 7.9(b). Numerical solutions for r ands obtained after
Chapter 7. 171
0 2 4 6 8 10 12 14 16 1810
−1
100
101
102
Number of iterations
Obje
ctiv
eFunct
ion
Fλ
(a)
0 5 10 15 2010
−2
10−1
100
101
Number of iterations
Root
Mea
nSquare
Err
or
(b)
0 0.05 0.1 0.15 0.2 0.25 0.31
1.5
2
2.5
t
r(t
)
(c)
0 0.5 1 1.5 2 2.5 3 3.5−4
−3
−2
−1
0
1
2
3
4
x
s(x
)
(d)
Figure 7.10: (a) The objective functionFλ, (b) the RMSEs forr(t) (− − −) ands(x) (· · · ) obtained using the hybrid-order regularisation (7.26) with regularisationparameterλopt = 2 for noise levelǫ0 = 0.1, and the numerical results(− · −) for (c)r(t) and (d)s(x) obtained using the minimisation process after 17 (unfixed) iterations,for Example 2. The corresponding analytical solutions (7.30) are shown by continuousline (—–) in (c) and (d).
27 (unfixed) and 21 (fixed) iterations are displayed in Figures 7.9(c) and 7.9(d), re-
spectively. As expected, the conclusions from Figure 7.9 obtained for a low level of
noiseǫ0 = 0.01 are very much the same as the those from Figure 7.8 obtained for no
noiseǫ0 = 0. From both Figures 7.8(c), 7.8(d) and 7.9(c), 7.9(d) one canobserve that
the numerical results are accurate and stable. Furthermore, there is little difference in
the results obtained whether we stop (fix) the iteration process at the minimum of the
RMSEs shown in Figures 7.8(b) and 7.9(b) or, if we let the iteration process running
Chapter 7. 172
(unfix) until converge of the regularised objective function is achieved.
Next, we consider a large amount of noise, such asǫ0 = 0.1, included in (7.32)
and the numerical results are shown in Figure 7.10. First, one can observe from Figure
7.10(a) that the convergence of the objective function (7.26) is rapidly achieved within
17 iterations and the monotonic decreasing curve has a somewhat different shape than
that recorded in Figure 7.8(a) for no noiseǫ0 = 0 or in Figure 7.9(a) for a low amount
noiseǫ0 = 0.01. Also, interestingly, unlike in Figures 7.8(b) and 7.9(b) where the
RMSEs show a minimum before the iteration process has finished, in Figure 7.10(b)
no such minimum occurs. Therefore, in Figures 7.10(c) and 7.10(d) we present only
numerical results forr ands, respectively, obtained after 17 (unfixed) iterations with
λopt = 2. From these figures it can be seen that the numerical solutions are stable, with
an unexpected very high accuracy in predicting thes component in Figure 7.10(d). For
completeness and clarity the RMSEs of Figures 7.7(b)–7.10(b) are given in numbers
in Table 7.3. From this table, and also from Figure 7.10(b), it can be seen that for
ǫ0 = 0.1 the components(x) is predicted more accurately than ther(t) component,
whilst the prediction forǫ0 ∈ 0, 0.01 is reversed.
Finally, we report that the numerical results presented in this example are compa-
rable in terms of accuracy and stability with the numerical results obtained recently
in [48] using a different method of successive approximantspreviously developed in
[47].
7.5.3 Example 3
The previous examples possessed an analytical (smooth) solution available explicitly
and they were tested in order to verify the accuracy and stability of the numerical
method employed. In this subsection, we consider a severe test example represented
Chapter 7. 173
by the non-smooth source components
r(t) =
t, 0 ≤ t ≤ 1/2,
1− t, 1/2 < t ≤ 1 = T,
(7.33)
s(x) =
x, 0 ≤ x ≤ 1/20,
0.1− x, 1/20 < x ≤ 1/10 = L,
(7.34)
whereL = 1/10, T = 1, X0 = 1/20. We also take the homogeneous initial tem-
perature (7.17). This example does not have an analytical solution for the temperature
u(x, t) readily available. Therefore, in such a situation the data (7.4) and (7.5) is sim-
ulated numerically by solving the direct problem (7.1) withthe multiplicative source
given by the product of the functions in (7.33) and (7.34), subject to the homogeneous
boundary and initial conditions (7.3) and (7.17). The BEM numerical solutions for the
dataχ(t) = u(0.2, t) andβ(x) = u(x, 1) are shown in Figure 7.11 for various numbers
of boundary elementsN = N0 ∈ 10, 20, 40. From this figure the convergence of the
numerical solution, as the number of boundary elements increases, can be observed.
Next we consider the inverse problem given by equations (7.1), (7.3), (7.6) with
0 0.2 0.4 0.6 0.8 10
1
2
3
4
5
6
7x 10
−3
t
χ(t
)
(a)
0 0.02 0.04 0.06 0.08 0.16.3
6.35
6.4
6.45
6.5
6.55
6.6x 10
−3
x
β(x
)
(b)
Figure 7.11: The numerical results for (a)χ(t) and (b)β(x) obtained using the BEMfor the direct problem withN = N0 ∈ 10(−·−), 20(· · · ), 40(−−−), for Example3.
Chapter 7. 174
0 1 2 3 4 5 6 7 810
−7
10−6
10−5
10−4
10−3
10−2
Number of iterations
Obje
ctiv
eFunct
ion
Fλ
(a)
0 0.2 0.4 0.6 0.8 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
t
r(t
)
(b)
0 0.02 0.04 0.06 0.08 0.10
0.01
0.02
0.03
0.04
0.05
0.06
x
s(x
)
(c)
Figure 7.12: (a) The objective functionFλ and the numerical results for (b)r(t) and (c)s(x) obtained with the hybrid-order regularisation (7.26) withregularisation parameterλ = 2 × 10−4 for p ∈ 1(− · −), 5(· · · ), 10(− − −)% noisy data for Example 3.The corresponding analytical solutions (7.33) and (7.34) are shown by continuous line(—–) in (b) and (c).
S0 = s(1/20) = 1/20 specified, (7.17), and the additional measured data (7.4) and
(7.5) which has been simulated numerically in Figure 7.11. We pick from Figure 7.11
the numerical BEM solutions obtained withN = N0 = 20 and we further perturb this
data with noise, as in (7.28). We took as initial guesses r¯0
= s¯0
= 0, and we initiated
the iterative minimisation process of the hybrid-order regularisation functional (7.26),
as described in Example 1. The numerical results obtained withλopt = 2×10−4 (found
by trial and error) are shown in Figure 7.12 forp ∈ 1, 5, 10% noise generated as in
(7.28). From Figure 7.12(a) it can be seen that the convergence of the functional (7.26)
Chapter 7. 175
is rapidly achieved within 7-8 iterations using thelsqnonlinroutine with TolFun = TolX
= 10−6. Also, Figures 7.12(b) and 7.12(c) show that stable and accurate numerical
solutions for bothr(t) ands(x) are obtained for all the amounts of noisep.
In closure, although not illustrated, we report that the same good performance has
been recorded when attempting to reconstruct even discontinuous source components.
7.6 Conclusions
In this chapter, inverse source problems with homogeneous Neumann boundary con-
ditions together with specified interior and final time temperature measurements have
been considered to find the space- and the time-dependent components of a multiplica-
tive source function. The numerical discretisation was based on the BEM combined
with a nonlinear Tikhonov regularisation procedure via thelsqnonlinroutine from the
MATLAB. For a wide range of test examples, the obtained results indicate that stable
and accurate numerical solutions have been achieved. The identification of both multi-
plicativer(t)s(x) and additiver1(t)+s1(x) components of space- and time-dependent
sources of the formr(t)s(x) + r1(t) + s1(x) can also be considered, [47], but its nu-
merical implementation is deferred to a future work.
Chapter 8. 176
Chapter 8
General Conclusions and Future Work
8.1 Conclusions
The aim of this thesis was to solve various inverse source problems for the (one-
dimensional) heat equation by using the BEM to find the time-dependent heat source
function, presented in Chapters 2–5, and space- and time-dependent heat source func-
tions for additive and multiplicative cases, presented in Chapters 6 and 7, respectively.
Several types of conditions such as non-local, non-classical, periodic, fixed point, time-
average and integral have been considered as boundary or overdetermination condi-
tions.
The BEM has been used as the main numerical approach for discretising the linear
heat equation with a heat source present. In Chapter 1 we havedescribed the BEM for
discretising the heat equation. With this method, the heat equation is first multiplied
by the fundamental solution and then integrated with the assistance of Green’s identity.
This leads to a boundary integral equation which can be discretised with resulting inte-
gral coefficients that can be evaluated analytically. The initial and boundary conditions
are also imposed.
In an inverse problem, additional conditions are required to determine uniquely the
unknown functions. However, this information has to come from measurements which
are contaminated with noise unavoidably. If the problem is ill-posed then small errors
177
Chapter 8. 178
in the measurement data result in highly and unbounded output solutions. Therefore,
regularisation methods need to be employed to deal with thisinstability.
In this thesis, many regularisation methods have been utilised together with the
BEM. A popular regularisation method, the Tikhonov regularisation, has been used
with orders zero, one and two. Additionally, for comparison, the TSVD method has
also been considered in Chapters 3 and 6. Moreover, in Chapter 4, the smoothing spline
technique has been considered as a regularisation method for seeking a regularised
first-order derivative of a noisy function.
Regularisation methods require a proper choice of the regularisation parameter.
There are many methods such as theL-curve method, the GCV criterion, and the
discrepancy principle which are all popular and successfulmethods for choosing the
regularisation parameter. TheL-curve method is the simplest method for choosing the
regularisation parameter. This method suggests choosing the parameter at the corner
of the L-curve which is a plot of the solution norm versus the corresponding resid-
ual for many positive regularisation parameters. Alternatively, we have also used the
GCV criterion in order to indicate a regularisation parameter, this method is based on
the minimising the GCV function of various positive regularisation parameters. When
the amount of noise is known, the discrepancy principle was proposed to be another
method for choosing the regularisation parameter. This method is more rigorous since
it requires the knowledge of the noise level with which the input data is contami-
nated. Furthermore, in Chapter 6, the selection of two regularisation parameters has
been based on theL-surface method, this method is a natural extension of theL-curve
method used for the selection of a single regularisation parameter. For comparison, the
simple trial and error technique has also been employed, i.e. various regularisation pa-
rameters were tested with gradually increasing value untiloscillations in the numerical
solutions have been stabilised.
To test the accuracy and stability of the BEM combined with regularisation meth-
ods, numerical examples consisting of various cases of unknown functions, such as
smooth continuous, non-smooth continuous and discontinuous functions, have been
Chapter 8. 179
illustrated and compared with their analytical solutions,where available. Otherwise,
in the cases where the analytical solution for the temperature is not available, we have
used the numerical solution of the corresponding direct problem and set the mesh size
to be different in the corresponding inverse problem presented in Chapters 3 and 5.
This is in order to avoid committing an inverse crime, see [32].
In summary, all numerical results with/without noise contamination have been
found to be accurate and stable. In Chapter 2, the determination of the time-dependent
heat source function and the temperature subjected to threegeneral boundary con-
ditions has been considered. These three conditions have been distinguished to be
six separate cases of boundary and overdetermination conditions and generating six
inverse problems. Some cases were found to be ill-conditioned; then the Tikhonov
regularisation with orders zero, one and two have been used on both exact and noisy
data. Whereas other cases were found to be well-conditionedand the use of BEM has
processed well for the inverse problem with no use of regularisation for the exact data,
but the regularisation was still needed when noise was present.
In Chapter 3, the identification of the time-dependent heat source and the temper-
ature subjected to a periodic boundary condition, a Robin boundary condition and an
integral overdetermination condition has been considered. The BEM has been devel-
oped and combined with two regularisation methods; the Tikhonov regularisation and
the TSVD method. A couple of benchmark test examples have been presented in order
to illustrate the accuracy of the numerical results. No regularisation was required in the
case of exact data and we found that the least-squares procedure and the SVD method
produced the same accurate numerical results. When noise was added, theL-curve
method and the discrepancy principle were selected for the appropriate choice of the
regularisation parameter, when using the Tikhonov regularisation of orders zero, one
and two, and the truncation level, when using the TSVD. Numerical results obtained
by using the BEM combined with either the TSVD or the zeroth-order Tikhonov regu-
larisation have been formed similar. The higher-order regularisation for smooth source
recovery gave more accurate results than the lower-order ones, while for non-smooth
Chapter 8. 180
sources the conclusion was reversed.
In Chapter 4, we have investigated the reconstruction of thetime-dependent blood
perfusion coefficient and the temperature in the bioheat equation subjected to the same
boundary and overdetermination conditions as in Chapter 3.A simple transformation
reduced the bioheat equation to be the classical heat equation, but now the overdeter-
mination condition contained the unknown source function.Two numerical examples
have been solved using the BEM. One example has been considered together with the
Tikhonov regularisation combined with the higher-order (of accuracy) finite difference
and use the GCV method as the choice of regularisation parameter. The second exam-
ple has used the BEM together with a smoothing spline technique for differentiating a
noisy function with apriori and aposteriorichoices of the regularisation parameters.
Chapter 5 presented an identification of the time-dependentheat source and the
temperature for the heat equation subjected to the non-classical boundary and integral
overdetermination conditions. We have utilised the same technique as before based
on the BEM together with the Tikhonov regularisation method. Three benchmark test
examples have been considered with smooth and non-smooth continuous source func-
tions to illustrate the accuracy and stability of the numerical results. Utilising the GCV
method as choice of regularisation parameter has performedwell to obtain stable and
accurate solution in all the investigated examples. We havealso found that there was
not much significant difference value of the regularisationparameter given by the dis-
crepancy principle or the trial and error technique.
In Chapter 6, we have investigated the more challenging identification of two un-
known source functions; the time- and space-dependent components of an additive
heat source and the temperature in the one-dimensional heatequation subjected to in-
terior point and time integral overdetermination conditions. The BEM was combined
with either the Tikhonov regularisation or the TSVD to solvethe inverse problem with
various selections of the regularisation parameter and thetruncation level based on the
L-curve method, the discrepancy principle and the GCV criterion when a single reg-
ularisation parameter was considered. We have also extended the analysis to the case
Chapter 8. 181
when two regularisation parameters were present and chose these parameters by using
theL-surface method.
The final case of inverse heat source problem presented in Chapter 7 was a non-
linear case study. This consisted of the simultaneous determination of multiplicative
space- and time-dependent source components and the temperature for the heat equa-
tion subject to homogeneous Neumann boundary condition, specified interior, and final
time temperature measurements. The numerical discretisation was based on the BEM
combined with a Tikhonov regularisation procedure. The resulting nonlinear optimi-
sation problem was solved using the MATLAB routinelsqnonlin. The hybrid-order
combination of the first- and second-order Tikhonov regularisation has achieved a sta-
ble and accurate numerical solution.
Throughout the thesis, the retrieved numerical results were found to be accurate and
stable concluding the reliability of the BEM combined with the various regularisation
techniques for solving a wide range of inverse source problems for the heat equation.
8.2 Future work
As we have studied so far, the use of the BEM combined with regularisation methods
can be developed for solving inverse source problems for theheat equation under var-
ious types of boundary and overdetermination conditions. This supports the idea that
the BEM combined with regularisation methods can also perform well in other related
possible future work, as follows.
(i) An inverse source problem related to that of Chapter 5 andgiven by the follow-
Chapter 8. 182
ing system of equations:
ut = uxx + r(t)f(x, t), (x, t) ∈ DT ,
u(x, 0) = u0(x), x ∈ [0, 1],
u(0, t) = 0, t ∈ (0, T ],
ut(1, t) + ux(1, t) + ϕ(u(1, t)) = 0, t ∈ (0, T ],
∫ 1
0u(x, t) dx = E(t), t ∈ [0, T ],
(8.1)
whereϕ is a given nonlinear function, has recently been investigated in [50] but with
no numerical study. It would be interesting to study the numerical reconstruction of
the time-depending heat sourcer(t) and the temperatureu(x, t) satisfying this inverse
problem by using the BEM together with the regularisation methods presented in this
thesis.
(ii) Another possible future work for the one-dimensional study is the combination
of the identifications in Chapters 6 and 7 for a more general heat source containing
both additive and multiplicative components, see [47]. This work is an identification
of finding the time-dependent source functionsr(t), r1(t), the space-dependent source
functionss(x), s1(x) and the temperatureu(x, t) which satisfy the heat equation
ut(x, t) = uxx(x, t) + r(t)s(x) + r1(t) + s1(x), (x, t) ∈ DT , (8.2)
subject to the initial condition (1.7), the homogeneous Neumann boundary condition
(7.3), and the additive measurements
u(X0, t) = χ(t), u(X1, t) = χ1(t), t ∈ [0, T ], (8.3)
u(x, T1) = β1(x), u(x, T2) = β2(x), x ∈ [0, L], (8.4)
at the fixed sensor locations0 ≤ X0 < X1 ≤ L and the fixed times0 < T1 < T2 ≤ T .
Chapter 8. 183
Fixing conditions are also required as
r(T1) = γ1, r(T2) = γ2, s(X0) = S0, s1(X0) = S0, (8.5)
where the functionsχ, χ1, µ1, µ2 and the constantsγ1, γ2, S0, S0 are given. This in-
verse problem is very challenging because it is nonlinear study and the Matlab routine
lsqnonlin will be required.
(iii) The multi-dimensional inverse source problem for theheat equation is also
very interesting to study further with the BEM. The following inverse source problem
can be further studied, see Cannon [8] and Yamamoto [62]. LetΩ be a bounded do-
main inRn, n = 1, 2, 3. Then, one can consider the inverse problem of finding the
temperatureu(x, t) for (x, t) ∈ Ω × (0, T ) and the space-dependent heat sourcef(x)
for x ∈ Ω, satisfying the transient heat conduction equation
∂u
∂t(x, t) = ∇2u(x, t) + r(t)f(x), (x, t) ∈ Ω× (0, T ), (8.6)
subject to the initial condition
u(x, 0) = u0(x), x ∈ Ω, (8.7)
and the overspecified Cauchy boundary data
u(x, t) = β(x, t), (x, t) ∈ ∂Ω× (0, T ), (8.8a)
∂u
∂n(x, t) = ϑ(x, t), (x, t) ∈ Γ× (0, T ), (8.8b)
or
∂u
∂n(x, t) = ϑ(x, t), (x, t) ∈ ∂Ω × (0, T ), (8.9a)
u(x, t) = β(x, t), (x, t) ∈ Γ× (0, T ), (8.9b)
Chapter 8. 184
whereΓ ⊂ ∂Ω is a non-empty open subset of the boundary∂Ω, andr, u0, β andϑ are
known functions.
Bibliography
[1] Ahmadabadi, M. N., Arab, M., and Maalek-Ghaini, F. (2009). The method of
fundamental solutions for the inverse space-dependent heat source problem.Engi-
neering Analysis with Boundary Elements, 33, 1231–1235.
[2] Beilina, L., Thanh, N. T., Klibanov, M. V., and Fiddy, M. A. (2014). Reconstruc-
tion from blind experimental data for an inverse problem fora hyperbolic equation.
Inverse Problems, 30, 025002 (24 pages).
[3] Belge, M., Kilmer, M., and Miller, E. L. (2002). Efficientdetermination of multiple
regularization parameters in a generalizedL-curve framework.Inverse Problems,
18, 1161–1183.
[4] Bourgeois, L., Chaulet, N., and Haddar, H. (2011). Stable reconstruction of gen-
eralized impedance boundary conditions.Inverse Problems, 27, 095002 (26 pages).
[5] Bourgeois, L., Chaulet, N., and Haddar, H. (2012). On simultaneous identification
of a scatterer and its generalized impedance boundary conditions.SIAM Journal on
Scientific Computing, 34, 1824–1848.
[6] Bourgeois, L. and Haddar, H. (2010). Identification of generalized impedance
boundary conditions in inverse scattering problems.Inverse Problems and Imaging,
4, 19–38.
[7] Cakoni, F. and Kress, R. (2013). Integral equation method for the inverse obstacle
185
Bibliography 186
problem with generalized impedance boundary condition.Inverse Problems, 29,
015005 (19 pages).
[8] Cannon, J. R. (1968). Determination of an unknown heat source from overspeci-
fied boundary data.SIAM Journal on Numerical Analysis, 5, 275–286.
[9] Cannon, J. R. (1984).The One-dimensional Heat Equation. Addison-Wesley
Publishing Company, Menlo Park, California.
[10] Cannon, J. R. and Hoek, J. (1986). Diffusion subject to the specification of mass.
Journal of Mathematical Analysis and Applications, 115, 517–529.
[11] Cannon, J. R., Lin, Y., and Wang, S. (1991). Determination of a control parameter
in a parabolic partial differential equation.Journal of the Australian Mathematical
Society, Series B, 33, 149–163.
[12] Chen, Z., Lu, Y., Xu, Y., and Yang, H. (2008). Multi-parameter Tikhonov regu-
larization for linear ill-posed operator equations.Journal of Computational Mathe-
matics, 26, 37–55.
[13] Elden, L., Berntsson, F., and Reginska, T. (2000). Wavelet and Fourier methods
for solving the sideways heat equation.SIAM Journal on Scientific Computing, 21,
2187–2205.
[14] Engl, H. W. and Kugler, P. (2005). Nonlinear inverse problem: Theoretical as-
pects and some industrial applications. In Capasso, V. and Periaux, J., editors,Mul-
tidisciplinary Methods for Analysis Optimization and Control of Complex Systems,
volume 6 ofthe series Mathematics in Industry, pages 3–47. Springer, Berlin.
[15] Farcas, A. and Lesnic, D. (2006). The boundary element method for the de-
termination of a heat source dependent on one variable.Journal of Engineering
Mathematics, 54, 375–388.
Bibliography 187
[16] Hansen, P. C. (2001). TheL-curve and its use in the numerical treatment of
inverse problems. In Johnston, P., editor,Computational Inverse Problems in Elec-
trocardiology, pages 119–142. WIT Press, Southampton.
[17] Hasanov, A. (2007). Simultaneous determination of source terms in a linear
parabolic problem from the final overdetermination: Weak solution approach.Jour-
nal of Mathermatical Analysis with Applications, 330, 766–779.
[18] Hazanee, A., Ismailov, M. I., Lesnic, D., and Kerimov, N. B. (2013). An inverse
time-dependent source problem for the heat equation.Applied Numerical Mathe-
matics, 69, 13–33.
[19] Hazanee, A., Ismailov, M. I., Lesnic, D., and Kerimov, N. B. (2015). An inverse
time-dependent source problem for the heat equation with a non-classical boundary
condition.Applied Mathematical Modelling, in press (available online).
[20] Hazanee, A. and Lesnic, D. (2013). Reconstruction of anadditive space- and
time-dependent heat source.European Journal of Computational Mechanics, 22,
304–329.
[21] Hazanee, A. and Lesnic, D. (2014). A time-dependent coefficient identification
problem for the bioheat equation. In Szczygiel, I., Nowak, A., and Rojczyk, M., ed-
itors,ICIPE2014 8th International Conference on Inverse Problems in Engineering,
pages 127–136. Silesian University of Technology, May 12-15, 2014, Poland.
[22] Ionkin, N. I. (1977). Solution of a boundary-value problem in heat conduction
with a non-classical boundary condition.Differential Equations, 13, 294–304.
[23] Ismailov, M. I. and Kanca, F. (2011). An inverse coefficient problem for a
parabolic equation in the case of nonlocal boundary and overdetermination con-
ditions. Mathematical Methods in the Applied Sciences, 34, 692–702.
Bibliography 188
[24] Ismailov, M. I., Kanca, F., and Lesnic, D. (2011). Determination of a time-
dependent heat source under nonlocal boundary and integraloverdetermination con-
ditions. Applied Mathematics and Computation, 218, 4138–4146.
[25] Ivanchov, M. (2003).Inverse Problems for Equations of Parabolic Type. VNTL
Publishers, Lviv.
[26] Ivanchov, M. I. (2001). Inverse problem for a multidimensional heat equation
with an unknown source function.Matematychni Studii, 16, 93–98.
[27] Ivanchov, N. I. (1993). Inverse problems for the heat-conduction equation with
nonlocal boundary conditions.Ukrainian Mathematical Journal, 45, 1186–1192.
[28] Ivanchov, N. I. (1995). On the determination of unknownsource in the heat
equation with nonlocal boundary conditions.Ukrainian Mathematics Journal, 47,
1647–1652.
[29] Johansson, B. T. and Lesnic, D. (2007). A variational method for identifying a
spacewise dependent heat source.IMA Journal of Applied Mathematics, 72, 748–
760.
[30] Johansson, B. T. and Lesnic, D. (2008). A procedure for determining a spacewise
dependent heat source and the initial temperature.Applicable Analysis, 87, 265–
276.
[31] Kabanikhin, S. I. (2008). Definitions and examples of inverse and ill-posed prob-
lems.Journal of Inverse and Ill-posed Problems, 16, 317–357.
[32] Kaipio, J. and Somersalo, E. (2007). Statistical inverse problems: Discretiza-
tion, model reduction and inverse crimes.Journal of Computational and Applied
Mathematics, 198, 493–504.
[33] Kerimov, N. B. and Ismailov, M. I. (2012). An inverse coefficient problem for the
heat equation in the case of nonlocal boundary conditions.Journal of Mathematical
Analysis and Applications, 396, 546–554.
Bibliography 189
[34] Kerimov, N. B. and Ismailov, M. I. (2015). Direct and inverse problems for the
heat equation with a dynamic type boundary condition.IMA Journal of Applied
Mathematics, accepted.
[35] Kuzhuget, A. V., Beilina, L., Klibanov, M., Sullivan, A., Nguyen, L., and Fiddy,
M. A. (2000). Blind experimental data collected in the field and an approximately
globally convergent inverse algorithm.Inverse Problems, 28, 095007 (33 pages).
[36] Langer, R. E. (1932). A problem in diffusion or in the flowof heat for a solid in
contact with a fluid.Tohoku Mathematical Journal, 35, 360–375.
[37] Lesnic, D. (2009). Identification of the time-dependent perfusion coefficient in
the bio-heat conduction equation.Journal of Inverse and Ill-posed Problems, 17,
753–764.
[38] Lesnic, D., Elliott, L., and Ingham, D. B. (1998). The boundary element so-
lution of the Laplace and biharmonic equations subjected tonoisy boundary data.
International Journal for Numerical Methods in Engineering, 43, 479–492.
[39] Lesnic, D., Elliott, L., Ingham, D. B., Knipe, R. J., andClennell, B. (1996). The
identification of hydraulic conductivities of composite rocks. In Jones, G., Fisher,
Q., and Knipe, R. J., editors,Faulting, Fault Sealing and Fluid Flow in Hydrocarbon
Reservoirs: Abstracts, pages 113–114.
[40] Morozov, A. (1966). On the solution of functional equations by the method of
regularization.Soviet Mathematics Doklady, 7, 414–417.
[41] Nakhushev, A. M. (1995).Equations of Mathematical Biology. Vysshaya Shkola,
Moscow, (in Russian).
[42] Petrov, Y. P. and Sizikov, V. S. (2005).Well-posed, Ill-posed and Intermediate
Problems. VSP, The Netherlands.
Bibliography 190
[43] Prilepko, A. I., Orlovsky, D. G., and Vasin, I. A. (2000). Methods for Solving
Inverse Problems in Mathematical Physics. M. Dekker, New York.
[44] Prilepko, A. I. and Solov’ev, V. V. (1988). Solvabilitytheorems and Rothe’s
method for inverse problems for a parabolic equation. I.Differential Equations, 23,
1230–1237.
[45] Prilepko, A. I. and Tkachenko, D. S. (2003). Inverse problem for a parabolic
equation with integral overdetermination.Journal of Inverse and Ill-Posed Prob-
lems, 11, 191–218.
[46] Samarskii, A. A. (1980). Some problems in differentialequations theory.Differ-
ential Equations, 16, 1221–1228.
[47] Savateev, E. G. (1995). On problems of determining the source function in a
parabolic equation.Journal of Inverse and Ill-Posed Problems, 3, 83-102.
[48] Shi, C., Wang, C., and Wei, T. (2014). Numerical solution for an inverse heat
source problem by an iterative method.Applied Mathematics and Computation,
244, 577–597.
[49] Skerget, P. and Brebbia, C. A. (1985). Time dependent non-linear potential prob-
lems. In Brebbia, C. A., editor,Boundary Element Research, pages 63–86. Springer-
Verlag, Berlin.
[50] Slodicka, M. (2015). A parabolic inverse source problem with a dynamic bound-
ary condition.Applied Mathematics and Computation, 256, 529–539.
[51] Slodicka, M., Lesnic, D., and Onyango, T. T. M. (2010). Determination of a time-
dependent heat transfer coefficient in a nonlinear inverse heat conduction problem.
Inverse Problems in Science and Engineering, 18, 65–81.
[52] Thanh, N. T., Beilina, L., Klibanov, M., and Fiddy, M. A.(2014). Reconstruction
of the refractive index from experimental backscattering data using a globally con-
vergent inverse method.SIAM Journal on Scientific Computing, 36, B273–B293.
Bibliography 191
[53] Trucu, D. (2009). Inverse Problems for Blood Perfusion Identification. PhD
Thesis, University of Leeds.
[54] Trucu, D., Ingham, D. B., and Lesnic, D. (2008). Inversetime-dependent perfu-
sion coefficient identification.Journal of Physics: Conference Series, 124, 012050
(28 pages).
[55] Trucu, D., Ingham, D. B., and Lesnic, D. (2010). Inversetemperature-dependent
perfusion coefficient reconstruction.International Journal of Non-Linear Mechan-
ics, 45, 542–549.
[56] Trucu, D., Ingham, D. B., and Lesnic, D. (2011). Reconstruction of the space-
and time-dependent blood perfusion coefficient in bio-heattransfer.Heat Transfer
Engineering, 32, 800–810.
[57] Twomey, S. (1963). On the numerical solution of Fredholm integral equations of
the first kind by the inversion of the linear system produced by quadrature.Journal
of the Association for Computing Machinery, 10, 97–101.
[58] Vogel, C. R. (1996). Non-convergence of theL-curve regularization parameter
selection method.Inverse Problems, 12, 535–547.
[59] Wei, T. and Li, M. (2006). High order numerical derivatives for one-dimensional
scattered noisy data.Applied Mathematics and Computation, 175, 1744–1759.
[60] Wrobel, L. C. (2002).The Boundary Element Method: Applications in Thermo-
Fluids and Acoustics,Vol.1. John Wiley & Sons, New York.
[61] Xiong, X., Yan, Y., and Wang, J. (2011). A direct numerical method for solving
inverse heat source problems.Journal of Physics: Conference Series, 290, 012017
(10 pages).
[62] Yamamoto, M. (1993). Conditional stability in determination of force terms of
heat equations in a rectangle.Mathematical and Computer Modelling, 18, 79–88.
Bibliography 192
[63] Yan, L., Fu, C. L., and Yang, F. L. (2008). The method of fundamental solutions
for the inverse heat source problem.Engineering Analysis with Boundary Elements,
32, 216–222.
[64] Yan, L., Yang, F. L., and Fu, C. L. (2009). A meshless method for solving an in-
verse spacewise-dependent heat source problem.Journal of Computational Physics,
228, 123–136.
[65] Yan, L., Yang, F. L., and Fu, C. L. (2010). A new numericalmethod for the
inverse source problem from a Bayesian perspective.International Journal for Nu-
merical Methods in Engineering, 85, 1460–1474.
[66] Yang, F. and Fu, C. L. (2010). A simplified Tikhonov regularization method for
determining the heat source.Applied Mathematical Modelling, 34, 3286–3299.
[67] Yang, L., Dehghan, M., Yu, J. N., and Luo, G. W. (2011). Inverse problem of
time-dependent heat sources numerical reconstruction.Mathematics and Comput-
ers in Simulation, 81, 1656–1672.
[68] Yi, Z. and Murio, D. A. (2004a). Source term identification in 1-D IHCP.Com-
puters and Mathematics with Applications, 47, 1921–1933.
[69] Yi, Z. and Murio, D. A. (2004b). Source term identification in 2-D IHCP.Com-
puters and Mathematics with Applications, 47, 1517–1533.